Volume 12, Number 12 December 2018 - University of Kent

8
Volume 12, Number 12 December 2018 thereasoner.org ISSN 1757-0522 Contents Editorial 93 Features 93 News 96 What’s Hot in . . . 96 Events 99 Courses and Programmes 99 Jobs and Studentships 100 Editorial Dear Readers, we’re close to a well deserved break, but of course reasoning never sleeps! It is my great plea- sure to introduce you to my interview with Juergen Lan- des. A mathematician who doesn’t mind at all span- ning philosophical founda- tions and practical applica- tions of inductive logic, Juer- gen touches on a number of topics of interest to the rea- soning community, and to the wider community of re- searchers. Many thanks to Juergen for sharing his thoughts and Seasonal Greetings to all Reasoners! Hykel Hosni University of Milan Features Interview with Juergen Landes Hykel Hosni : Can you tell us about your background and how you eventually got interested in Reasoning? Juergen Landes: As a kid, my dream was to become a re- searcher making discoveries. And since the corporate world isn’t much interested in discoveries for the sake of discovery, it was clear to me that I had to attend a university. I read popu- lar science books on the universe and Grand Unification The- ory (which aims to provide a theory which unifies all physical 93

Transcript of Volume 12, Number 12 December 2018 - University of Kent

Volume 12 Number 12December 2018

thereasonerorgISSN 1757-0522

Contents

Editorial 93

Features 93

News 96

Whatrsquos Hot in 96

Events 99

Courses and Programmes 99

Jobs and Studentships 100

EditorialDear Readers wersquore closeto a well deserved breakbut of course reasoning neversleeps It is my great plea-sure to introduce you to myinterview with Juergen Lan-des A mathematician whodoesnrsquot mind at all span-ning philosophical founda-tions and practical applica-tions of inductive logic Juer-gen touches on a number oftopics of interest to the rea-soning community and tothe wider community of re-searchers Many thanks to

Juergen for sharing his thoughts and Seasonal Greetings to allReasoners

Hykel HosniUniversity of Milan

Features

Interview with Juergen LandesHykel Hosni Can you tell us about your background and howyou eventually got interested in ReasoningJuergen Landes As a kid my dream was to become a re-searcher making discoveries And since the corporate worldisnrsquot much interested in discoveries for the sake of discovery itwas clear to me that I had to attend a university I read popu-lar science books on the universe and Grand Unification The-ory (which aims to provide a theory which unifies all physical

93

forces) Scientific American and Physik in unserer Zeit Even-tually I enrolled in my home town (Frankfurt) with two majorsin the subjects which interested me most in school maths andphysics

I soon dropped physics for two reasons i) I liked theoret-ical physics much more than experimental physics but hatedthe idea that a great theory I might invent could be refuted byNature No such thing happens to mathematical theories ii)Continuing to study physics would have entailed conductingclassical experiments in the afternoons on instruments designedfor students to conduct exactly these experiments making errorcalculations and writing reports This was wasting my time Ifelt So I decided to become a mathematician instead

My main interests were in the most fundamental questionsthere are in maths so I specialised in logic I thus moved toFreiburg where I wrote my Diploma Thesis (which roughlymeans a combined bachelor and master thesis in todayrsquos terms)in Parametrized Complexity Theory This theory aims to pro-vide answers to the question How long does it take an algo-rithm to solve instances of a class of problems For examplegiven a finite list of integers what is the maximum run time ofan algorithm A to correctly order the numbers by size

After completing my studies I wanted to do a PhD inlogic Without a possibility to continue in Freiburg I entereda crowded job market After a frustratingly long time I finallyreceived an offer from Jeff Paris in Manchester to work on un-certain reasoning which I gladly accepted That is how I fi-nally became interested in reasoning proper I guess we arejust not born with innate interest in reasoning What a shame

HH What was your PhD thesis onJL I worked on the Principle of Spectrum Exchangeability

in polyadic Pure Inductive LogicLet me explain Pure In-

ductive Logic ldquois the study ofrational probability treatedas a branch of mathematicallogicrdquo (2015 Pure InductiveLogic Paris Jeff B amp Ven-covska Alena CUP) Induc-tive Logic proper originateswith Rudolf Carnap and hisfamous Continuum of Induc-tive Methods It starts outas an exercise in explicatingcommon sense uncertain reasoning it is a formal approach tocapture how rational agents reason in the face of uncertaintyTechnically the goal is to assign probabilities to sentences ofa logical language The probabilities are interpreted as theagentrsquos degree of belief in the sentences being true

The main methodology for assigning probabilities is to in-vestigate principles (axioms) which aim to capture (parts of)human rationality These axioms put down constraints on theprobabilities one may assign With axioms and a formal lan-guage in place the formal mathematical analysis motivated byphilosophical and practical considerations can be brought tobear

Typically one aims to show that all probability functions sat-isfying a set of axioms can be represented as integrals over par-ticularly simple probability functions Such results are knownas de Finetti-style representation theorems first proved for ex-changeable sequences (1974 Theory of Probability Bruno deFinetti Wiley) Such a representation theorem not only gives

one a better view of all probability functions consistent witha set of axioms it also often allows for much simpler proofsfirst one shows that every simple probability function in therepresentation theorem has property P then one shows that anyconvex combination of simple functions with property P alsopossesses property P hence the axioms entail P QED

Before I started to work on Pure Inductive Logic most workwas on first order languages with unary relation symbols with-out equality and function symbols At first I investigated anaxiom of exchangeability (Spectrum Exchangeability) on firstorder languages with relation symbols of arbitrary arity andwithout equality and function symbols I later added the equal-ity symbol to the language which brought to light interestingconnections to the principle of Language Invariance (2011 Asurvey of some recent results on Spectrum Exchangeability inPolyadic Inductive Logic Landes Jurgen and Paris Jeff B andVencovska Alena Synthese 19-47) This highly intuitive prin-ciple requires that inferences about a particular matter at handought not to change if the underlying language is enriched byfurther (relation) symbols about which we have zero informa-tion

The Manchester group has continued to work on polyadiclanguages (eg PhD Theses by Elizabeth Howarth and TahelRonel and (Paris amp Vencovska 2015 Pure Inductive LogicCUP))

HH And thenJL Since leaving Manchester in 2009 Irsquove been a postdoc

on four different projects working on uncertain reasoning invarying contexts

My first day in regular employment happened to be my 30thbirthday My role was to support the Life Cycle Analysis(LCA) of micro algae producing methane by modelling multi-criterial decision making under uncertainty LCA aims to listand assess the environmental impacts that goods and servicesproduce Unfortunately the PI of the project was diagnosedwith cancer half-way through this one-year project ndash hersquos madea full recovery This left me as the only formal person in a labof actual scientists getting their hands dirty with methanisationprocesses By now these methanisation processes have madetheir way into the real world you may find them in biogas in-stallations at farms near you

For the second postdoc I moved to Munich where I workedon designing automated contract negotiations in temporary em-ployment agencies via multi-agent systems

My third postdoc was with Jon Williamson in Kent (youmight know him as founder of The Reasoner) who happensto be a former student of Jeff Paris While this job was backon familiar formal grounds it was my first job in philosophyWe worked on uncertain reasoning in the objective Bayesianapproach Roughly this approach works as follows one firstdetermines the set of probability functions (E) consistent withthe evidence On first pass one may think itrsquos reasonable toadopt any function in E However this approach requires one topick the function in E which has greatest entropy ( if there is aunique such function) We showed how to justify this approachaxiomatically (2013 Objective Bayesianism and the maximumentropy principle Jurgen Landes and Jon Williamson Entropy3528-3591) [my most cited paper] and (2015 Justifying objec-tive Bayesianism on predicate languages 17(4) 2459-2543)

HH Do you think then that MaxEnt is the best justified prin-ciple of uncertain reasoning

JL While this maximum entropy approach is language in-

94

variant if relation symbols are added one has zero informationabout It fails to be language invariant if we add formal expres-sions for chances to the language It fails what we call ChanceInvariance Under this construal of invariant inferences it isthe Centre of Mass approach which simply picks out the cen-tre of mass of the acceptable probability functions which sat-isfies Chance Invariance Against all hope I think this peculiarfinding should get more press (2017 Invariant EquivocationJurgen Landes and George Masterton Erkenntnis 141-167)

Irsquove just completed my fourth postdoc On this projectwe aimed to assign a probability to the causal claim that adrug causes a particular adverse drug reaction While thereare many approaches marrying probabilities and causation weused the Bayesian network machinery to ldquomerelyrdquo assign aprobability to one causal hypothesis (2018 Epistemology ofCausal Inference in Pharmacology Jurgen Landes Barbara Os-imani Roland Poellinger European Journal for Philosophy ofScience 3-49) Irsquove been much interested in the confirma-tory value of varied evidence to the causal hypothesis in thisframework (2019 Variety of Evidence Jurgen Landes Erken-ntnis forthcoming DOI 101007s10670-018-0024-6 moremanuscripts under review)

HH Whatrsquos next thenJL Early this month Irsquove returned from paternity leave af-

ter the birth of our third child in August 2018 I now work onmy own research project on Evidence and Objective BayesianEpistemology In this three-year project I will investigate justi-fications and applications of amalgamating all the available ev-idence within the objective Bayesian approach Why and howdoes one (best) aggregate the available evidence

HH Thatrsquos a fascinating and timely question ndash we look for-ward to reading about your take on the problem To more per-sonal questions now You certainly crossed borders betweencountries disciplines and sectors over the past decade To-day the rise of rightwing populists and anti-EU nationalists arepushing forward a view which will make it increasingly harderfor the new generation of scientists to do that Any thoughts

JL Yes way too many thoughtsFirst we all donrsquot walk a mile in someone elsersquos shoes anymore before we judge In fact we donrsquot even seem to considera different opinion upbringing or socio-economic backgroundother than our own to be proper As a result societies are be-coming more and more partitioned into ever more divided sec-tions in which opinions from the outside no further penetrateThis facilitates populism around the world

Take us academics for example more and more of us (areforced to) work in different countries where we work at uni-versities in large cities and communicate with colleagues andincreasingly also staff in English all day every day Every-one has a decently waged also not necessarily stable workcontract We engage in academic cerebral exercises which areworlds apart from a typical day of early school leavers withoutany prospects for a decent job We all are living in our ownbubbles although we are ever more connected over the Inter-net

Irsquom deeply troubled by any nationalist movement indepen-dent of country and time We should always remember thatthere may be a time in the much too near future in which thebombs might be falling (the economy collapse natural disasterstrike etc) right where we stand now Wouldnrsquot it then be greatto have some friends who could shelter our kids To me thatrsquosa no-brainer

The best and clinching argument for the EU must be the longlong years of peace it has brought to Europe Just remembernot so long ago France and England fought a war that lasted for100 years Preventing wars for decades must be worth the tinybit of autonomy politicians (not countries) give up as well asall membership fees

As for crossing disciplines I can only advise to limit thenumber of cross-overs to a small number Every time you crossyou start from close to zero you lack a grasp of the interest-ing debates an understanding of the type of argument requiredto publish in top journals a personal network a well-stackedlibrary and so on Eventually you may well end up with a pub-lication record that does not impress in university departmentssince too many publications are outside their sphere of inter-est (2013 Embedding philosophers in the practices of sciencebringing humanities to the sciences Nancy Tuana Synthese1955-1973) We welcome back the above-mentioned bubbles

HH Can you see drawbacks in crossing bordersJL Crossing countries is a somewhat different matter That

is something I found deeply rewarding Furthermore it madevery clear to me that every country is populated by human be-ings which are not so different from each other Therersquos abso-lutely no reason for countries to get angry at another Itrsquos longoverdue that we all grew upHowever there are serious downsides to being an academic no-mad Irsquom not talking about the inevitable minor issues likecontributions to different pension systems hassle with differ-ent bureaucracies and trips to embassies Irsquom talking about theproblem of making friends and maintaining friendships over along period of time Just imagine every time you meet an in-teresting person you immediately think in a few yearrsquos timewe will live in different cities and probably different countriesThatrsquos not idealThe problem is worse for those of us who have children Rais-ing the young is extremely time consuming and tiring Nothaving a social support network in place (parentsgood friendsliving around the corner) because of a recent move does notmake things any easier Been there done it sent a postcard (3times)

HH Speaking of populisms we desperately need to getmore people interested in sound reasoning and argumentationYou said before itrsquos a shame it is not an innate interest of oursso can you think of actions we could take as a community toplay a leading role in urging the general public to pay attentionto reasoning

JL Most of our work is far far removed from the every-daylives of ordinary people I donrsquot think that we stand a seriouschance in engaging them in our academic brain activities

Grabbing the attention of inquisitive and excitable childrenwho may well recognise the value of good reasoning appearsmore effective in the long run to me Irsquove recently attended atalk by Andy Oxman He and his colleagues teach kids aboutrecognising and evaluating claims Scaling up this and othersuch initiatives is a way forward I believe in

While I think that getting the general public interested inreasoning as too big a task involving people facing compli-cated problems routinely at work makes more sense to me Thework by Vincenzo Crupi and others (eg 2017 Understandingand improving decisions in clinical medicine (I) Reasoningheuristics and error Internal and Emergency Medicine Vin-cenzo Crupi and Fabrizio Elia 689-691) is the type of project Iwould like to see more of

95

Finally engaging the interested public at a Ted Conferencesay is a good idea

News

Calls for PapersKnowing the Unknown Philosophical Perspectives on Igno-rance special issue of Synthese deadline 20 FebruaryHybrid Data and Knowledge Driven Decision Making underUncertainty special issue of Information Sciencesl deadline30 FebruaryThought Experiments in theHistory of Philosophy of Sciencespecial issue of HOPOS deadline 31 MarchFolk Psychology Pluralistic Approaches special issue ofSynthese deadline 15 MayImprecise Probabilities Logic and Rationality special issueof International Journal of Approximate Reasoning deadline 1June

Whatrsquos Hot in

Medieval ReasoningMany times students andcolleagues alike have askedme what would be a good in-troduction to medieval logicand I always struggle to finda good answer Usually Ibegin by telling them wherenot to start from Williamand Martha Kneale The De-velopment of Logic (Oxford1962) (And not becauseKneale amp Kneale is not aclassic groundbreaking textndash it is both Nor because it hasnrsquot aged gracefully ndash although ithasnrsquot aged particularly well at all But because one would getthe impression that medieval logicians ndash despite reaching stun-ning levels of apparent sophistication ndash got some very simplestuff very very wrong I donrsquot think that this is a fair evaluationnor is it a historically correct assessment Overall itrsquos better toleave Kneale amp Kneale on the side at the beginning perhaps tobe revisited later on) However pointing beginners in the direc-tion of a good introductory textbook is a trickier business Mostoverviews of medieval logic are either too partial or are fairlyobsolete ndash quite often both ndash and in many cases not really fitfor readers lacking a background in Ancient and Medieval phi-losophy Alexander Broadiersquos Introduction to Medieval Logic(2nd ed Oxford 1993) while being a fairly accessible and ba-sic text is a bit outdated and despite its title focuses heavily onthe first half of the 14th century Just as elementary (albeit justas partial and outdated as well) Paul Vincent Spadersquos ThoughtsWords and Things An Introduction to Late Medieval Logicand Semantic Theory (version 11 2002 httpspvspadecomLogicdocsthoughts1_1apdf) has both the faultsand virtues of an informal handbook put together from lecturenotes and circulated but never polished for publication JanPinborgrsquos Medieval Semantics Selected Studies on MedievalLogic and Grammar (English translation London 1984) while

not as basic has more thematic and chronologic breadth butunfortunately its age shows Normann Kretzmannrsquos AnthonyKennyrsquos and Jan Pinborgrsquos (eds) The Cambridge History ofLater Medieval Philosophy (Cambridge 1982) is largely de-voted to medieval logic and semantics to the point that itrsquos stillone of the most complete overviews around Some more re-cent overviews that collect a good number of rich and detailedarticles are Dov Gabbayrsquos and John Woodsrsquo (eds) Handbookof the History of Logic II Mediaeval and Renaissance Logic(Amsterdam 2008) and Catarina Dutilh Novaesrsquo and StephenReadrsquos (eds) Cambridge Companion to Medieval Logic (Cam-bridge 2016) Some other studies and anthologies come tomind ndash among those that are at least partially in English KlausJacobirsquos (ed) Argumentationstheorie (1993) Mikko Yrjonsu-urirsquos (ed) Medieval Formal Logic (2001) Dutilh Novaesrsquo For-malizing Medieval Logical Theories (2007) and Terry ParsonrsquosArticulating Medieval Logic (2014)

The above are all very interesting reads but ndash for the mostpart ndash they donrsquot seem to make for viable introductions It lookslike a Medieval Logic for Dummies (from 400 to 1450 ca) hasyet to be written ndash and until somebody does I wonrsquot be able togive a short simple answer to those who ask In the meantimetherersquos plenty of exciting stuff to skim through if you want todive into the subject headfirst

Graziana CiolaUCLA

Uncertain ReasoningTwenty years ago this De-cember James M Joycepublished ldquoA NonpragmaticVindication of Probabilismrdquo(Philosophy of Science1998) This paper inspireda whole new way to thinkabout epistemic norms Thebasic idea of rdquoscoring rulesrdquoor rdquoinaccuracy measuresrdquowas not new to Joyce in-deed the basics have beenaround since at least thefifties If yoursquore predicting whether or not it will rain and yourpredictions are given in the form of ldquoyesrdquo ldquonordquo answers itrsquosclear what counts as a good forecast if it rains you did wellif you said that it would and you did poorly if you said that itwouldnrsquot If instead you give your predictions in the form ofa probability of rain itrsquos less clear how to measure success Ifit rains and you said that the probability of rain was 90 thatis quite good ndash better than saying 80 worse than saying 95ndash but how exactly should one measure the quality of forecastsThis is important since we use probabilistic forecasts in ourdecision making and so it is important to know how to assessthe quality of a forecast In essence what one wants is someform of rdquodistance from the truthrdquo which one could use tomeasure how ldquocloserdquo your probabilistic predictions were Butthere are a great many different ways to measure the distancefrom the truth how can we choose between them

What Joyce pointed out was that while it might be unclearwhich measure of forecast quality is the one we ought to usethere are some features that they arguably ought to share andthat those common features are enough to build an argument

96

that our degrees of belief ought to obey the axioms of proba-bility theory (among other things) The details of Joycersquos orig-inal argument neednrsquot concern us here there have been manydifferent versions of the general approach since 1998 The ar-gumentative strategy ndash argue that certain principles of scoringrules constrain whatever epistemic utility function governs ourepistemic behaviour use these principles to prove what epis-temic norms we ought to obey ndash is one that has been takenup by many people and put to many uses Not just proba-bilism but conditionalisation and the principal principle havebeen justified in broadly this way This way of proceeding ar-gues Joyce is one that avoids the pitfalls and problems thatappear to plague the other standard way of justifying normsof credence namely ldquosure lossrdquo betting arguments (What aretypically called ldquoDutch book argumentsrdquo but for reasons I men-tioned last month Irsquom avoiding that term) For an up to dateoverview of the literature inspired by Joycersquos paper see RichardPettigrewrsquos recent book ldquoAccuracy and the Laws of Credencerdquo(OUP 2016)

This continues to be an interesting and fruitful area of re-search and Irsquoll return to the topic in a future column

Seamus BradleyPhilosophy University of Tilburg

Mathematical Philosophy

The current spectaculardevelopments in machinelearning go hand in handwith a growing popularinterest in the discipline Butinterest of formal philoso-phers in machine learningpredates the disciplinersquosrecent rise from an obscurebranch of artificial intelli-gence to its epitome activityThe reason is that machinelearning and formal epistemology share a concern with thesame fundamental question how to learn from data How toinfer conclusions from given data conclusions that go beyondthe data itself How could this be formalized or evenmdashthecharacteristic business of machine learningmdashautomatized

Thus several approaches within machine learning have beendirected by epistemologists at philosophical problems In someinstances like causal reasoning (Pearl Causality Cambridge2009) there has long been much interaction between philoso-phers and computer scientists As another example formallearning theory initiated by among others Putnam (J SymbLog 30(1) 49ndash57 1965) evolved into a technical field in com-puter science (Osherson et al Systems That Learn MIT 1985)and has more recently been developed further as a projectwithin philosophy of science (Kelly The Logic of ReliableInquiry Oxford 1996) Other instances are more unidirec-tional Schurz (Philos Sci 75 278ndash305 2008) has proposeda Reichenbachian justification of induction based on results incompetitive online learning (Cesa-Bianchi and Lugosi Predic-tion Learning and Games Cambridge 2006) Harman andKulkarni (Reliable Reasoning MIT 2006) have highlightedthe philosophical relevance of statistical learning theory theframework that underlies supervised learning (Vapnik Statisti-

cal Learning Theory Wiley 1998)On a more grand level a number of philosophers have argued

that machine learning prompts novel ways of doing philosophyof science This is in itself not new Thagard (MIT 1988) forinstance defended a computational philosophy of science in-spired by artificial intelligence Nor is this the norm Harmanand Kulkarni employ the statistical learning theory frameworkin a largely traditional analysis of the questions of inductionand its justification like the Goodman riddle and the role ofsimplicity But work within formal learning theory is explic-itly committed to a more pragmatic means-ends epistemology(Schulte Brit J Philos Sci 50(1) 1ndash31 1999) that insteadof the traditional focus on justifying inductive reasoning in gen-eral starts with the complexity of a given inductive problemthus determining the epistemic goals that are still feasible forit Wheeler (The Routledge Companion to Philosophy of So-cial Science 321ndash329 2017) goes much further still and de-clares ldquotraditional Enlightenment epistemologyrdquo bankrupt tobe replaced by the ldquoprimacy of practicerdquo of a pragmatist ma-chine epistemology

Certainly the concurrent advent of big data has contributedto the perception that machine learning is largely a practicalenterprise a matter of unleashing rough-and-ready algorithmson vast amounts of data In the most extreme version of thispicture theoretical analysis is not just hopelessly lagging be-hind it is superfluous and some have even speculated noto-riously about the consequences for scientific theorizing Thispresumed opposition between theory and practice however isalso a theme within machine learning Vapnik in the intro-ductory chapter to The Nature of Statistical Learning Theory(2nd edition Springer 2000) paints a picture of the history ofmachine learning where the practitioner is set against the theo-rist Vapnik chides the ldquoartificial intelligence hardlinersrdquo (ldquotheywho declared that lsquoComplex theories do not work simple al-gorithms dorsquordquo) for repeatedly making excessive promises andconcludes confidently that ldquoa new methodological situation inthe learning problem has developed where practical methodsare the result of a deep theoretical analysis of the statisticalbounds rather than the result of inventing new smart heuris-ticsrdquo Written still before the modern rehabilitation of neuralnetworks in the form of deep learning this now sounds overlyconfident The theme has indeed recently flared up again inthe community following Rahimi and Rechtrsquos speech at NIPS2017 in which they liken current machine learning research toalchemy

The predictable outrage notwithstanding Rahimi and Rechtnowhere commit to theory for the sake of theory they do nottake issue with the primacy of the pragmatic goal of findingalgorithms that work well in practice Their argument is thata theoretical understanding is important for this goal it is thestep back that facilitates so to speak a more efficient searchthrough the space of possible algorithms Continuous with thistheoretical work a further step back would be the foundationalwork that is the province of philosophers This concerns forinstance what it should mean when Rahimi and Recht speakabout ldquorigorous reliable verifiable knowledgerdquo and how thiscould be grounded in the relevant mathematics This also con-cerns the outer limits on learning algorithmsrsquo capabilities thesubject of my own work on universal prediction (University ofGroningen 2018)

The gap between theory and practice is particularly conspic-uous in what has been called a paradox of deep learning neural

97

networks perform astonishingly well in practice much betterthan they should in theory This is illustrated by Zhang et al(ldquoUnderstanding deep learning requires rethinking generaliza-tionrdquo ICLR 2017) with a simple experiment essentially theauthors took some standard architectures and data sets and thenchanged the data by reshuffling the labels in a random mannerThis guaranteed that there was no relation between instancesand labels yet the networks still managed to attain extremelylow training error entailing a wide divergence between train-ing error and generalization error (the latter must still be veryhigh the labeling is fully random so there is nothing to learn)But explanations based on statistical learning theory rely on aconnection between training error (or more generally the archi-tecture of the network) and generalization error wherefore thedemonstrated possibility of varying the latter while keeping theformer constant implies that these explanations simply do notapply Depending on onersquos perspective one could take this asanother illustration of the impotence of theory or of the direneed for a greater theoretical effort In either case the properway to understand deep learning is an important ongoing de-bate in the machine learning community that merits attentionfrom philosophers too

Tom SterkenburgMunich Centre for Mathematical Philosophy

Evidence-Based Medicine

Precision medicine is not new but it is currently a hot topic inEBM Often used interchangeably with personalised medicinethis new approach to healthcare aims to individualise treat-ments by recognizing the ways in which patients can differ per-son to person in the nature of disease and in patient responseto treatments It is well known that patients will respond dif-ferently to treatments that are effective for other people withthe same disease It is being increasingly recognised that thisis due to biological differences in the patients and the diseasesthemselves often at the genetic level Two very recent papersarticulate what precision medicine aims to do identify the ge-netic differences among patients with a certain disease and usenew technologies to enable us to make sense of the level ofcomplexity associated with using genetics to make predictions

One study classified subtypes of acute myeloid leukemia(AML) by their gene expression and DNA-methylation pat-terns AML is a single disease as it has a specific characteri-zation ldquoblocked myeloid lineage differentiation and accumu-lation of leukemic blast cellsrdquo But it is a ldquohighly heterogeneousdiseaserdquo as it can be caused by many different types of geneticalterations These different genetic alterations define separatesub-types of AML and patients with different sub-types willrequire different specific treatments This study identified pat-terns in the expression of genes in tumor cells from AML pa-tients which revealed gene regulatory networks that are spe-cific to the different sub-types of AML There were a numberof recurrent alterations important for lukemic growth identifiedthat present possibilities for targeted treatment Without thisknowledge patients may receive treatment that will not be ef-fective as it is not targeted against their specific AML subtype

The results of this study are claimed to be able to base tar-geted treatments for specific AML sub-types It is importantto note here that while precision medicine aims to take resultssuch as these and use them to target individual patients it is

actually sub-populations of patients with specific AML sub-types that can be targeted rather than individual patients per seDrugs will not be tailored to any one individual but novel drugscan be designed with specific sub-populations in mind and ex-tant drugs can be better targeted at specific sub-populationsThis is why precision medicine may be a better term than per-sonalised medicine as these techniques allow us to be moreprecise in targeting treatments to any one individual This mayseem like a small semantic difference but being clear about thistempers the ability to make grandiose claims about treating in-dividuals - it still might be the case that precision treatmentsare not effective for individuals due to chance or other factorsboth biological and environmental

Difficulties face the implementation of this research bothbroadly and specifically A difficulty facing precision medicineis practical in nature the level of complexity associated withusing results from disciplines such as genetics One way inwhich precision medicine hopes to overcome this challenge isthrough the use of technology in particular AI A practical dif-ficulty specific to oncology is that ldquotumors change over timeprogressing from benign to malignant becoming metastaticand developing resistance to certain therapiesrdquo Essentiallythe tumor cells evolve resulting in lsquointratumour heterogeneityrsquoThis means that due to tumour evolution using genetic mark-ers of cancer subtypes may quickly not be useful for targetingtherapies An enterprise hoping to provide solutions to bothdifficulties is the REVOLVER system that has recently pub-lished results on predicting changes in tumour evolution usingan AI system This system uses a machine-learning approachknown as transfer learning (TL)19 to predict the next steps inspecific tumor cell evolution Tumor evolution seems to be pre-dictable due to the prognostic value of histopathological stag-ing and molecular markers Problems facing using this data topredict tumor evolution are the complexity of patterns in theavailable data and stochastic forces in the mechanisms of tu-mor cell evolution The REVOLVER system can deal betterwith complexity and stochasticity than a human reasoner Thisenables more precise information for specific cancer patients asto what might happen during the course of their disease

I have written about AI in this column before and with thatcase (triaging in opthamology) the system was again used towork with large quantities of data to improve the efficiency ofhealth care delivery Precision medicine offers much in the wayof improving how we treat severe diseases but as the data setsget larger and more complex we will increasingly have to turnto AI It is promising that there is development in this fieldthat can meet this challenge Interestingly at present the datasets are not actually large enough for the AI systems to providethe high degree of precision we require But with precisionmedicine research considered a major growth area in medicinewe will likely not have to wait too long to overcome this prob-lem

DJ Auker-HowlettPhilosophy Kent

98

Events

December

KEiMS Conference on Knowledge Exchange in the Mathe-matical Sciences Aston University 3ndash4 DecemberCvC Causation vs Constitution Loosening the FrictionBergen 3ndash4 DecemberAxiMet Axiomatizing MetatheorymdashTruth Provability andBeyond Salzburg Austria 6ndash7 DecemberML4H Workshop on Machine Learning for Health MontrealCanada 8 DecemberDiI Workshop on Disagreement in Inquiry University of Tartu8 DecemberRLPO Reinforcement Learning Under Partial ObservabilityMontreal Canada 8 DecemberRPiB Real Patterns University of Barcelona 14 DecemberWrsquosBaD Whatrsquos so Bad About Dialetheism Kyoto Univer-sity Japan 15ndash17 December

January

PoMaL Graduate Conference on the Philosophy of Mathemat-ics and Logic University of Cambridge 19ndash20 JanuaryFiSS Foundations in Social SciencemdashMechanisms ActionsFunctions University of Duisburg-Essen Germany 24ndash25 Jan-uaryVoUiS Varieties of Understanding in Science Utrecht TheNetherlands 24ndash25 January

February

DMaMG Dark Matter and Modified Gravity ConferenceAachen Germany 6ndash8 February

Courses and Programmes

Courses

SSA Summer School on Argumentation Computational andLinguistic Perspectives on Argumentation Warsaw Poland 6ndash10 September

Programmes

APhil MAPhD in Analytic Philosophy University ofBarcelonaMaster Programme MA in Pure and Applied Logic Univer-sity of BarcelonaDoctoral Programme in Philosophy Language Mind andPractice Department of Philosophy University of ZurichSwitzerlandDoctoral Programme in Philosophy Department of Philoso-phy University of Milan ItalyLogiCS Joint doctoral program on Logical Methods in Com-puter Science TU Wien TU Graz and JKU Linz AustriaHPSM MA in the History and Philosophy of Science andMedicine Durham UniversityMaster Programme in Statistics University College DublinLoPhiSC Master in Logic Philosophy of Science and Epis-temology Pantheon-Sorbonne University (Paris 1) and Paris-Sorbonne University (Paris 4)Master Programme in Artificial Intelligence Radboud Uni-versity Nijmegen the NetherlandsMaster Programme Philosophy and Economics Institute ofPhilosophy University of BayreuthMA in Cognitive Science School of Politics InternationalStudies and Philosophy Queenrsquos University BelfastMA in Logic and the Philosophy ofMathematics Departmentof Philosophy University of BristolMA Programmes in Philosophy of Science University ofLeedsMA in Logic and Philosophy of Science Faculty of PhilosophyPhilosophy of Science and Study of Religion LMU MunichMA in Logic and Theory of Science Department of Logic ofthe Eotvos Lorand University Budapest HungaryMA in Metaphysics Language and Mind Department of Phi-losophy University of LiverpoolMA inMind Brain and Learning Westminster Institute of Ed-ucation Oxford Brookes UniversityMA in Philosophy by research Tilburg UniversityMA in Philosophy Science and Society TiLPS Tilburg Uni-versityMA in Philosophy of Biological and Cognitive Sciences De-partment of Philosophy University of BristolMA in Rhetoric School of Journalism Media and Communi-cation University of Central LancashireMA programmes in Philosophy of Language and Linguisticsand Philosophy of Mind and Psychology University of Birm-inghamMRes in Methods and Practices of Philosophical ResearchNorthern Institute of Philosophy University of AberdeenMSc in Applied Statistics Department of Economics Mathe-matics and Statistics Birkbeck University of LondonMSc in Applied Statistics and Datamining School of Mathe-matics and Statistics University of St Andrews

99

MSc in Artificial Intelligence Faculty of Engineering Uni-versity of LeedsMSc in Cognitiveamp Decision Sciences Psychology UniversityCollege LondonMSc in Cognitive Systems Language Learning and Reason-ing University of PotsdamMSc in Cognitive Science University of Osnabruck GermanyMSc in Cognitive PsychologyNeuropsychology School ofPsychology University of KentMSc in Logic Institute for Logic Language and ComputationUniversity of AmsterdamMSc in Mind Language amp Embodied Cognition School ofPhilosophy Psychology and Language Sciences University ofEdinburghMSc in Philosophy of Science Technology and Society Uni-versity of Twente The NetherlandsMRes in Cognitive Science and Humanities Language Com-munication and Organization Institute for Logic CognitionLanguage and Information University of the Basque Country(Donostia San Sebastian)OpenMind International School of Advanced Studies in Cog-nitive Sciences University of BucharestResearchMaster in Philosophy and Economics Erasmus Uni-versity Rotterdam The Netherlands

Jobs and Studentships

JobsAssistant Professor in Epistemology California State Uni-versity at Sacramento deadline until filledAssistant Professor in Logic amp Epistemology University ofNorth Carolina at Greensboro deadline until filledAssistant Professor in Philosophy of Medicine University ofNorth Carolina at Greensboro deadline until filledAssistant Professor in Philosophy of Science University ofFlorida deadline until filledLecturer in Actuarial Science University of California SantaBarbara deadline until filledProfessor in Philosophy of Science City University of NewYork deadline 3 DecemberResearch Associate in Bayesian modelling University ofSheffield deadline 12 DecemberResearch Fellow in Data Science University of Bristol dead-line 3 JanuaryLecturer in Statistical Science University College Londondeadline 9 February

100

  • Editorial
  • Features
  • News
  • Whats Hot in hellip
  • Events
  • Courses and Programmes
  • Jobs and Studentships

forces) Scientific American and Physik in unserer Zeit Even-tually I enrolled in my home town (Frankfurt) with two majorsin the subjects which interested me most in school maths andphysics

I soon dropped physics for two reasons i) I liked theoret-ical physics much more than experimental physics but hatedthe idea that a great theory I might invent could be refuted byNature No such thing happens to mathematical theories ii)Continuing to study physics would have entailed conductingclassical experiments in the afternoons on instruments designedfor students to conduct exactly these experiments making errorcalculations and writing reports This was wasting my time Ifelt So I decided to become a mathematician instead

My main interests were in the most fundamental questionsthere are in maths so I specialised in logic I thus moved toFreiburg where I wrote my Diploma Thesis (which roughlymeans a combined bachelor and master thesis in todayrsquos terms)in Parametrized Complexity Theory This theory aims to pro-vide answers to the question How long does it take an algo-rithm to solve instances of a class of problems For examplegiven a finite list of integers what is the maximum run time ofan algorithm A to correctly order the numbers by size

After completing my studies I wanted to do a PhD inlogic Without a possibility to continue in Freiburg I entereda crowded job market After a frustratingly long time I finallyreceived an offer from Jeff Paris in Manchester to work on un-certain reasoning which I gladly accepted That is how I fi-nally became interested in reasoning proper I guess we arejust not born with innate interest in reasoning What a shame

HH What was your PhD thesis onJL I worked on the Principle of Spectrum Exchangeability

in polyadic Pure Inductive LogicLet me explain Pure In-

ductive Logic ldquois the study ofrational probability treatedas a branch of mathematicallogicrdquo (2015 Pure InductiveLogic Paris Jeff B amp Ven-covska Alena CUP) Induc-tive Logic proper originateswith Rudolf Carnap and hisfamous Continuum of Induc-tive Methods It starts outas an exercise in explicatingcommon sense uncertain reasoning it is a formal approach tocapture how rational agents reason in the face of uncertaintyTechnically the goal is to assign probabilities to sentences ofa logical language The probabilities are interpreted as theagentrsquos degree of belief in the sentences being true

The main methodology for assigning probabilities is to in-vestigate principles (axioms) which aim to capture (parts of)human rationality These axioms put down constraints on theprobabilities one may assign With axioms and a formal lan-guage in place the formal mathematical analysis motivated byphilosophical and practical considerations can be brought tobear

Typically one aims to show that all probability functions sat-isfying a set of axioms can be represented as integrals over par-ticularly simple probability functions Such results are knownas de Finetti-style representation theorems first proved for ex-changeable sequences (1974 Theory of Probability Bruno deFinetti Wiley) Such a representation theorem not only gives

one a better view of all probability functions consistent witha set of axioms it also often allows for much simpler proofsfirst one shows that every simple probability function in therepresentation theorem has property P then one shows that anyconvex combination of simple functions with property P alsopossesses property P hence the axioms entail P QED

Before I started to work on Pure Inductive Logic most workwas on first order languages with unary relation symbols with-out equality and function symbols At first I investigated anaxiom of exchangeability (Spectrum Exchangeability) on firstorder languages with relation symbols of arbitrary arity andwithout equality and function symbols I later added the equal-ity symbol to the language which brought to light interestingconnections to the principle of Language Invariance (2011 Asurvey of some recent results on Spectrum Exchangeability inPolyadic Inductive Logic Landes Jurgen and Paris Jeff B andVencovska Alena Synthese 19-47) This highly intuitive prin-ciple requires that inferences about a particular matter at handought not to change if the underlying language is enriched byfurther (relation) symbols about which we have zero informa-tion

The Manchester group has continued to work on polyadiclanguages (eg PhD Theses by Elizabeth Howarth and TahelRonel and (Paris amp Vencovska 2015 Pure Inductive LogicCUP))

HH And thenJL Since leaving Manchester in 2009 Irsquove been a postdoc

on four different projects working on uncertain reasoning invarying contexts

My first day in regular employment happened to be my 30thbirthday My role was to support the Life Cycle Analysis(LCA) of micro algae producing methane by modelling multi-criterial decision making under uncertainty LCA aims to listand assess the environmental impacts that goods and servicesproduce Unfortunately the PI of the project was diagnosedwith cancer half-way through this one-year project ndash hersquos madea full recovery This left me as the only formal person in a labof actual scientists getting their hands dirty with methanisationprocesses By now these methanisation processes have madetheir way into the real world you may find them in biogas in-stallations at farms near you

For the second postdoc I moved to Munich where I workedon designing automated contract negotiations in temporary em-ployment agencies via multi-agent systems

My third postdoc was with Jon Williamson in Kent (youmight know him as founder of The Reasoner) who happensto be a former student of Jeff Paris While this job was backon familiar formal grounds it was my first job in philosophyWe worked on uncertain reasoning in the objective Bayesianapproach Roughly this approach works as follows one firstdetermines the set of probability functions (E) consistent withthe evidence On first pass one may think itrsquos reasonable toadopt any function in E However this approach requires one topick the function in E which has greatest entropy ( if there is aunique such function) We showed how to justify this approachaxiomatically (2013 Objective Bayesianism and the maximumentropy principle Jurgen Landes and Jon Williamson Entropy3528-3591) [my most cited paper] and (2015 Justifying objec-tive Bayesianism on predicate languages 17(4) 2459-2543)

HH Do you think then that MaxEnt is the best justified prin-ciple of uncertain reasoning

JL While this maximum entropy approach is language in-

94

variant if relation symbols are added one has zero informationabout It fails to be language invariant if we add formal expres-sions for chances to the language It fails what we call ChanceInvariance Under this construal of invariant inferences it isthe Centre of Mass approach which simply picks out the cen-tre of mass of the acceptable probability functions which sat-isfies Chance Invariance Against all hope I think this peculiarfinding should get more press (2017 Invariant EquivocationJurgen Landes and George Masterton Erkenntnis 141-167)

Irsquove just completed my fourth postdoc On this projectwe aimed to assign a probability to the causal claim that adrug causes a particular adverse drug reaction While thereare many approaches marrying probabilities and causation weused the Bayesian network machinery to ldquomerelyrdquo assign aprobability to one causal hypothesis (2018 Epistemology ofCausal Inference in Pharmacology Jurgen Landes Barbara Os-imani Roland Poellinger European Journal for Philosophy ofScience 3-49) Irsquove been much interested in the confirma-tory value of varied evidence to the causal hypothesis in thisframework (2019 Variety of Evidence Jurgen Landes Erken-ntnis forthcoming DOI 101007s10670-018-0024-6 moremanuscripts under review)

HH Whatrsquos next thenJL Early this month Irsquove returned from paternity leave af-

ter the birth of our third child in August 2018 I now work onmy own research project on Evidence and Objective BayesianEpistemology In this three-year project I will investigate justi-fications and applications of amalgamating all the available ev-idence within the objective Bayesian approach Why and howdoes one (best) aggregate the available evidence

HH Thatrsquos a fascinating and timely question ndash we look for-ward to reading about your take on the problem To more per-sonal questions now You certainly crossed borders betweencountries disciplines and sectors over the past decade To-day the rise of rightwing populists and anti-EU nationalists arepushing forward a view which will make it increasingly harderfor the new generation of scientists to do that Any thoughts

JL Yes way too many thoughtsFirst we all donrsquot walk a mile in someone elsersquos shoes anymore before we judge In fact we donrsquot even seem to considera different opinion upbringing or socio-economic backgroundother than our own to be proper As a result societies are be-coming more and more partitioned into ever more divided sec-tions in which opinions from the outside no further penetrateThis facilitates populism around the world

Take us academics for example more and more of us (areforced to) work in different countries where we work at uni-versities in large cities and communicate with colleagues andincreasingly also staff in English all day every day Every-one has a decently waged also not necessarily stable workcontract We engage in academic cerebral exercises which areworlds apart from a typical day of early school leavers withoutany prospects for a decent job We all are living in our ownbubbles although we are ever more connected over the Inter-net

Irsquom deeply troubled by any nationalist movement indepen-dent of country and time We should always remember thatthere may be a time in the much too near future in which thebombs might be falling (the economy collapse natural disasterstrike etc) right where we stand now Wouldnrsquot it then be greatto have some friends who could shelter our kids To me thatrsquosa no-brainer

The best and clinching argument for the EU must be the longlong years of peace it has brought to Europe Just remembernot so long ago France and England fought a war that lasted for100 years Preventing wars for decades must be worth the tinybit of autonomy politicians (not countries) give up as well asall membership fees

As for crossing disciplines I can only advise to limit thenumber of cross-overs to a small number Every time you crossyou start from close to zero you lack a grasp of the interest-ing debates an understanding of the type of argument requiredto publish in top journals a personal network a well-stackedlibrary and so on Eventually you may well end up with a pub-lication record that does not impress in university departmentssince too many publications are outside their sphere of inter-est (2013 Embedding philosophers in the practices of sciencebringing humanities to the sciences Nancy Tuana Synthese1955-1973) We welcome back the above-mentioned bubbles

HH Can you see drawbacks in crossing bordersJL Crossing countries is a somewhat different matter That

is something I found deeply rewarding Furthermore it madevery clear to me that every country is populated by human be-ings which are not so different from each other Therersquos abso-lutely no reason for countries to get angry at another Itrsquos longoverdue that we all grew upHowever there are serious downsides to being an academic no-mad Irsquom not talking about the inevitable minor issues likecontributions to different pension systems hassle with differ-ent bureaucracies and trips to embassies Irsquom talking about theproblem of making friends and maintaining friendships over along period of time Just imagine every time you meet an in-teresting person you immediately think in a few yearrsquos timewe will live in different cities and probably different countriesThatrsquos not idealThe problem is worse for those of us who have children Rais-ing the young is extremely time consuming and tiring Nothaving a social support network in place (parentsgood friendsliving around the corner) because of a recent move does notmake things any easier Been there done it sent a postcard (3times)

HH Speaking of populisms we desperately need to getmore people interested in sound reasoning and argumentationYou said before itrsquos a shame it is not an innate interest of oursso can you think of actions we could take as a community toplay a leading role in urging the general public to pay attentionto reasoning

JL Most of our work is far far removed from the every-daylives of ordinary people I donrsquot think that we stand a seriouschance in engaging them in our academic brain activities

Grabbing the attention of inquisitive and excitable childrenwho may well recognise the value of good reasoning appearsmore effective in the long run to me Irsquove recently attended atalk by Andy Oxman He and his colleagues teach kids aboutrecognising and evaluating claims Scaling up this and othersuch initiatives is a way forward I believe in

While I think that getting the general public interested inreasoning as too big a task involving people facing compli-cated problems routinely at work makes more sense to me Thework by Vincenzo Crupi and others (eg 2017 Understandingand improving decisions in clinical medicine (I) Reasoningheuristics and error Internal and Emergency Medicine Vin-cenzo Crupi and Fabrizio Elia 689-691) is the type of project Iwould like to see more of

95

Finally engaging the interested public at a Ted Conferencesay is a good idea

News

Calls for PapersKnowing the Unknown Philosophical Perspectives on Igno-rance special issue of Synthese deadline 20 FebruaryHybrid Data and Knowledge Driven Decision Making underUncertainty special issue of Information Sciencesl deadline30 FebruaryThought Experiments in theHistory of Philosophy of Sciencespecial issue of HOPOS deadline 31 MarchFolk Psychology Pluralistic Approaches special issue ofSynthese deadline 15 MayImprecise Probabilities Logic and Rationality special issueof International Journal of Approximate Reasoning deadline 1June

Whatrsquos Hot in

Medieval ReasoningMany times students andcolleagues alike have askedme what would be a good in-troduction to medieval logicand I always struggle to finda good answer Usually Ibegin by telling them wherenot to start from Williamand Martha Kneale The De-velopment of Logic (Oxford1962) (And not becauseKneale amp Kneale is not aclassic groundbreaking textndash it is both Nor because it hasnrsquot aged gracefully ndash although ithasnrsquot aged particularly well at all But because one would getthe impression that medieval logicians ndash despite reaching stun-ning levels of apparent sophistication ndash got some very simplestuff very very wrong I donrsquot think that this is a fair evaluationnor is it a historically correct assessment Overall itrsquos better toleave Kneale amp Kneale on the side at the beginning perhaps tobe revisited later on) However pointing beginners in the direc-tion of a good introductory textbook is a trickier business Mostoverviews of medieval logic are either too partial or are fairlyobsolete ndash quite often both ndash and in many cases not really fitfor readers lacking a background in Ancient and Medieval phi-losophy Alexander Broadiersquos Introduction to Medieval Logic(2nd ed Oxford 1993) while being a fairly accessible and ba-sic text is a bit outdated and despite its title focuses heavily onthe first half of the 14th century Just as elementary (albeit justas partial and outdated as well) Paul Vincent Spadersquos ThoughtsWords and Things An Introduction to Late Medieval Logicand Semantic Theory (version 11 2002 httpspvspadecomLogicdocsthoughts1_1apdf) has both the faultsand virtues of an informal handbook put together from lecturenotes and circulated but never polished for publication JanPinborgrsquos Medieval Semantics Selected Studies on MedievalLogic and Grammar (English translation London 1984) while

not as basic has more thematic and chronologic breadth butunfortunately its age shows Normann Kretzmannrsquos AnthonyKennyrsquos and Jan Pinborgrsquos (eds) The Cambridge History ofLater Medieval Philosophy (Cambridge 1982) is largely de-voted to medieval logic and semantics to the point that itrsquos stillone of the most complete overviews around Some more re-cent overviews that collect a good number of rich and detailedarticles are Dov Gabbayrsquos and John Woodsrsquo (eds) Handbookof the History of Logic II Mediaeval and Renaissance Logic(Amsterdam 2008) and Catarina Dutilh Novaesrsquo and StephenReadrsquos (eds) Cambridge Companion to Medieval Logic (Cam-bridge 2016) Some other studies and anthologies come tomind ndash among those that are at least partially in English KlausJacobirsquos (ed) Argumentationstheorie (1993) Mikko Yrjonsu-urirsquos (ed) Medieval Formal Logic (2001) Dutilh Novaesrsquo For-malizing Medieval Logical Theories (2007) and Terry ParsonrsquosArticulating Medieval Logic (2014)

The above are all very interesting reads but ndash for the mostpart ndash they donrsquot seem to make for viable introductions It lookslike a Medieval Logic for Dummies (from 400 to 1450 ca) hasyet to be written ndash and until somebody does I wonrsquot be able togive a short simple answer to those who ask In the meantimetherersquos plenty of exciting stuff to skim through if you want todive into the subject headfirst

Graziana CiolaUCLA

Uncertain ReasoningTwenty years ago this De-cember James M Joycepublished ldquoA NonpragmaticVindication of Probabilismrdquo(Philosophy of Science1998) This paper inspireda whole new way to thinkabout epistemic norms Thebasic idea of rdquoscoring rulesrdquoor rdquoinaccuracy measuresrdquowas not new to Joyce in-deed the basics have beenaround since at least thefifties If yoursquore predicting whether or not it will rain and yourpredictions are given in the form of ldquoyesrdquo ldquonordquo answers itrsquosclear what counts as a good forecast if it rains you did wellif you said that it would and you did poorly if you said that itwouldnrsquot If instead you give your predictions in the form ofa probability of rain itrsquos less clear how to measure success Ifit rains and you said that the probability of rain was 90 thatis quite good ndash better than saying 80 worse than saying 95ndash but how exactly should one measure the quality of forecastsThis is important since we use probabilistic forecasts in ourdecision making and so it is important to know how to assessthe quality of a forecast In essence what one wants is someform of rdquodistance from the truthrdquo which one could use tomeasure how ldquocloserdquo your probabilistic predictions were Butthere are a great many different ways to measure the distancefrom the truth how can we choose between them

What Joyce pointed out was that while it might be unclearwhich measure of forecast quality is the one we ought to usethere are some features that they arguably ought to share andthat those common features are enough to build an argument

96

that our degrees of belief ought to obey the axioms of proba-bility theory (among other things) The details of Joycersquos orig-inal argument neednrsquot concern us here there have been manydifferent versions of the general approach since 1998 The ar-gumentative strategy ndash argue that certain principles of scoringrules constrain whatever epistemic utility function governs ourepistemic behaviour use these principles to prove what epis-temic norms we ought to obey ndash is one that has been takenup by many people and put to many uses Not just proba-bilism but conditionalisation and the principal principle havebeen justified in broadly this way This way of proceeding ar-gues Joyce is one that avoids the pitfalls and problems thatappear to plague the other standard way of justifying normsof credence namely ldquosure lossrdquo betting arguments (What aretypically called ldquoDutch book argumentsrdquo but for reasons I men-tioned last month Irsquom avoiding that term) For an up to dateoverview of the literature inspired by Joycersquos paper see RichardPettigrewrsquos recent book ldquoAccuracy and the Laws of Credencerdquo(OUP 2016)

This continues to be an interesting and fruitful area of re-search and Irsquoll return to the topic in a future column

Seamus BradleyPhilosophy University of Tilburg

Mathematical Philosophy

The current spectaculardevelopments in machinelearning go hand in handwith a growing popularinterest in the discipline Butinterest of formal philoso-phers in machine learningpredates the disciplinersquosrecent rise from an obscurebranch of artificial intelli-gence to its epitome activityThe reason is that machinelearning and formal epistemology share a concern with thesame fundamental question how to learn from data How toinfer conclusions from given data conclusions that go beyondthe data itself How could this be formalized or evenmdashthecharacteristic business of machine learningmdashautomatized

Thus several approaches within machine learning have beendirected by epistemologists at philosophical problems In someinstances like causal reasoning (Pearl Causality Cambridge2009) there has long been much interaction between philoso-phers and computer scientists As another example formallearning theory initiated by among others Putnam (J SymbLog 30(1) 49ndash57 1965) evolved into a technical field in com-puter science (Osherson et al Systems That Learn MIT 1985)and has more recently been developed further as a projectwithin philosophy of science (Kelly The Logic of ReliableInquiry Oxford 1996) Other instances are more unidirec-tional Schurz (Philos Sci 75 278ndash305 2008) has proposeda Reichenbachian justification of induction based on results incompetitive online learning (Cesa-Bianchi and Lugosi Predic-tion Learning and Games Cambridge 2006) Harman andKulkarni (Reliable Reasoning MIT 2006) have highlightedthe philosophical relevance of statistical learning theory theframework that underlies supervised learning (Vapnik Statisti-

cal Learning Theory Wiley 1998)On a more grand level a number of philosophers have argued

that machine learning prompts novel ways of doing philosophyof science This is in itself not new Thagard (MIT 1988) forinstance defended a computational philosophy of science in-spired by artificial intelligence Nor is this the norm Harmanand Kulkarni employ the statistical learning theory frameworkin a largely traditional analysis of the questions of inductionand its justification like the Goodman riddle and the role ofsimplicity But work within formal learning theory is explic-itly committed to a more pragmatic means-ends epistemology(Schulte Brit J Philos Sci 50(1) 1ndash31 1999) that insteadof the traditional focus on justifying inductive reasoning in gen-eral starts with the complexity of a given inductive problemthus determining the epistemic goals that are still feasible forit Wheeler (The Routledge Companion to Philosophy of So-cial Science 321ndash329 2017) goes much further still and de-clares ldquotraditional Enlightenment epistemologyrdquo bankrupt tobe replaced by the ldquoprimacy of practicerdquo of a pragmatist ma-chine epistemology

Certainly the concurrent advent of big data has contributedto the perception that machine learning is largely a practicalenterprise a matter of unleashing rough-and-ready algorithmson vast amounts of data In the most extreme version of thispicture theoretical analysis is not just hopelessly lagging be-hind it is superfluous and some have even speculated noto-riously about the consequences for scientific theorizing Thispresumed opposition between theory and practice however isalso a theme within machine learning Vapnik in the intro-ductory chapter to The Nature of Statistical Learning Theory(2nd edition Springer 2000) paints a picture of the history ofmachine learning where the practitioner is set against the theo-rist Vapnik chides the ldquoartificial intelligence hardlinersrdquo (ldquotheywho declared that lsquoComplex theories do not work simple al-gorithms dorsquordquo) for repeatedly making excessive promises andconcludes confidently that ldquoa new methodological situation inthe learning problem has developed where practical methodsare the result of a deep theoretical analysis of the statisticalbounds rather than the result of inventing new smart heuris-ticsrdquo Written still before the modern rehabilitation of neuralnetworks in the form of deep learning this now sounds overlyconfident The theme has indeed recently flared up again inthe community following Rahimi and Rechtrsquos speech at NIPS2017 in which they liken current machine learning research toalchemy

The predictable outrage notwithstanding Rahimi and Rechtnowhere commit to theory for the sake of theory they do nottake issue with the primacy of the pragmatic goal of findingalgorithms that work well in practice Their argument is thata theoretical understanding is important for this goal it is thestep back that facilitates so to speak a more efficient searchthrough the space of possible algorithms Continuous with thistheoretical work a further step back would be the foundationalwork that is the province of philosophers This concerns forinstance what it should mean when Rahimi and Recht speakabout ldquorigorous reliable verifiable knowledgerdquo and how thiscould be grounded in the relevant mathematics This also con-cerns the outer limits on learning algorithmsrsquo capabilities thesubject of my own work on universal prediction (University ofGroningen 2018)

The gap between theory and practice is particularly conspic-uous in what has been called a paradox of deep learning neural

97

networks perform astonishingly well in practice much betterthan they should in theory This is illustrated by Zhang et al(ldquoUnderstanding deep learning requires rethinking generaliza-tionrdquo ICLR 2017) with a simple experiment essentially theauthors took some standard architectures and data sets and thenchanged the data by reshuffling the labels in a random mannerThis guaranteed that there was no relation between instancesand labels yet the networks still managed to attain extremelylow training error entailing a wide divergence between train-ing error and generalization error (the latter must still be veryhigh the labeling is fully random so there is nothing to learn)But explanations based on statistical learning theory rely on aconnection between training error (or more generally the archi-tecture of the network) and generalization error wherefore thedemonstrated possibility of varying the latter while keeping theformer constant implies that these explanations simply do notapply Depending on onersquos perspective one could take this asanother illustration of the impotence of theory or of the direneed for a greater theoretical effort In either case the properway to understand deep learning is an important ongoing de-bate in the machine learning community that merits attentionfrom philosophers too

Tom SterkenburgMunich Centre for Mathematical Philosophy

Evidence-Based Medicine

Precision medicine is not new but it is currently a hot topic inEBM Often used interchangeably with personalised medicinethis new approach to healthcare aims to individualise treat-ments by recognizing the ways in which patients can differ per-son to person in the nature of disease and in patient responseto treatments It is well known that patients will respond dif-ferently to treatments that are effective for other people withthe same disease It is being increasingly recognised that thisis due to biological differences in the patients and the diseasesthemselves often at the genetic level Two very recent papersarticulate what precision medicine aims to do identify the ge-netic differences among patients with a certain disease and usenew technologies to enable us to make sense of the level ofcomplexity associated with using genetics to make predictions

One study classified subtypes of acute myeloid leukemia(AML) by their gene expression and DNA-methylation pat-terns AML is a single disease as it has a specific characteri-zation ldquoblocked myeloid lineage differentiation and accumu-lation of leukemic blast cellsrdquo But it is a ldquohighly heterogeneousdiseaserdquo as it can be caused by many different types of geneticalterations These different genetic alterations define separatesub-types of AML and patients with different sub-types willrequire different specific treatments This study identified pat-terns in the expression of genes in tumor cells from AML pa-tients which revealed gene regulatory networks that are spe-cific to the different sub-types of AML There were a numberof recurrent alterations important for lukemic growth identifiedthat present possibilities for targeted treatment Without thisknowledge patients may receive treatment that will not be ef-fective as it is not targeted against their specific AML subtype

The results of this study are claimed to be able to base tar-geted treatments for specific AML sub-types It is importantto note here that while precision medicine aims to take resultssuch as these and use them to target individual patients it is

actually sub-populations of patients with specific AML sub-types that can be targeted rather than individual patients per seDrugs will not be tailored to any one individual but novel drugscan be designed with specific sub-populations in mind and ex-tant drugs can be better targeted at specific sub-populationsThis is why precision medicine may be a better term than per-sonalised medicine as these techniques allow us to be moreprecise in targeting treatments to any one individual This mayseem like a small semantic difference but being clear about thistempers the ability to make grandiose claims about treating in-dividuals - it still might be the case that precision treatmentsare not effective for individuals due to chance or other factorsboth biological and environmental

Difficulties face the implementation of this research bothbroadly and specifically A difficulty facing precision medicineis practical in nature the level of complexity associated withusing results from disciplines such as genetics One way inwhich precision medicine hopes to overcome this challenge isthrough the use of technology in particular AI A practical dif-ficulty specific to oncology is that ldquotumors change over timeprogressing from benign to malignant becoming metastaticand developing resistance to certain therapiesrdquo Essentiallythe tumor cells evolve resulting in lsquointratumour heterogeneityrsquoThis means that due to tumour evolution using genetic mark-ers of cancer subtypes may quickly not be useful for targetingtherapies An enterprise hoping to provide solutions to bothdifficulties is the REVOLVER system that has recently pub-lished results on predicting changes in tumour evolution usingan AI system This system uses a machine-learning approachknown as transfer learning (TL)19 to predict the next steps inspecific tumor cell evolution Tumor evolution seems to be pre-dictable due to the prognostic value of histopathological stag-ing and molecular markers Problems facing using this data topredict tumor evolution are the complexity of patterns in theavailable data and stochastic forces in the mechanisms of tu-mor cell evolution The REVOLVER system can deal betterwith complexity and stochasticity than a human reasoner Thisenables more precise information for specific cancer patients asto what might happen during the course of their disease

I have written about AI in this column before and with thatcase (triaging in opthamology) the system was again used towork with large quantities of data to improve the efficiency ofhealth care delivery Precision medicine offers much in the wayof improving how we treat severe diseases but as the data setsget larger and more complex we will increasingly have to turnto AI It is promising that there is development in this fieldthat can meet this challenge Interestingly at present the datasets are not actually large enough for the AI systems to providethe high degree of precision we require But with precisionmedicine research considered a major growth area in medicinewe will likely not have to wait too long to overcome this prob-lem

DJ Auker-HowlettPhilosophy Kent

98

Events

December

KEiMS Conference on Knowledge Exchange in the Mathe-matical Sciences Aston University 3ndash4 DecemberCvC Causation vs Constitution Loosening the FrictionBergen 3ndash4 DecemberAxiMet Axiomatizing MetatheorymdashTruth Provability andBeyond Salzburg Austria 6ndash7 DecemberML4H Workshop on Machine Learning for Health MontrealCanada 8 DecemberDiI Workshop on Disagreement in Inquiry University of Tartu8 DecemberRLPO Reinforcement Learning Under Partial ObservabilityMontreal Canada 8 DecemberRPiB Real Patterns University of Barcelona 14 DecemberWrsquosBaD Whatrsquos so Bad About Dialetheism Kyoto Univer-sity Japan 15ndash17 December

January

PoMaL Graduate Conference on the Philosophy of Mathemat-ics and Logic University of Cambridge 19ndash20 JanuaryFiSS Foundations in Social SciencemdashMechanisms ActionsFunctions University of Duisburg-Essen Germany 24ndash25 Jan-uaryVoUiS Varieties of Understanding in Science Utrecht TheNetherlands 24ndash25 January

February

DMaMG Dark Matter and Modified Gravity ConferenceAachen Germany 6ndash8 February

Courses and Programmes

Courses

SSA Summer School on Argumentation Computational andLinguistic Perspectives on Argumentation Warsaw Poland 6ndash10 September

Programmes

APhil MAPhD in Analytic Philosophy University ofBarcelonaMaster Programme MA in Pure and Applied Logic Univer-sity of BarcelonaDoctoral Programme in Philosophy Language Mind andPractice Department of Philosophy University of ZurichSwitzerlandDoctoral Programme in Philosophy Department of Philoso-phy University of Milan ItalyLogiCS Joint doctoral program on Logical Methods in Com-puter Science TU Wien TU Graz and JKU Linz AustriaHPSM MA in the History and Philosophy of Science andMedicine Durham UniversityMaster Programme in Statistics University College DublinLoPhiSC Master in Logic Philosophy of Science and Epis-temology Pantheon-Sorbonne University (Paris 1) and Paris-Sorbonne University (Paris 4)Master Programme in Artificial Intelligence Radboud Uni-versity Nijmegen the NetherlandsMaster Programme Philosophy and Economics Institute ofPhilosophy University of BayreuthMA in Cognitive Science School of Politics InternationalStudies and Philosophy Queenrsquos University BelfastMA in Logic and the Philosophy ofMathematics Departmentof Philosophy University of BristolMA Programmes in Philosophy of Science University ofLeedsMA in Logic and Philosophy of Science Faculty of PhilosophyPhilosophy of Science and Study of Religion LMU MunichMA in Logic and Theory of Science Department of Logic ofthe Eotvos Lorand University Budapest HungaryMA in Metaphysics Language and Mind Department of Phi-losophy University of LiverpoolMA inMind Brain and Learning Westminster Institute of Ed-ucation Oxford Brookes UniversityMA in Philosophy by research Tilburg UniversityMA in Philosophy Science and Society TiLPS Tilburg Uni-versityMA in Philosophy of Biological and Cognitive Sciences De-partment of Philosophy University of BristolMA in Rhetoric School of Journalism Media and Communi-cation University of Central LancashireMA programmes in Philosophy of Language and Linguisticsand Philosophy of Mind and Psychology University of Birm-inghamMRes in Methods and Practices of Philosophical ResearchNorthern Institute of Philosophy University of AberdeenMSc in Applied Statistics Department of Economics Mathe-matics and Statistics Birkbeck University of LondonMSc in Applied Statistics and Datamining School of Mathe-matics and Statistics University of St Andrews

99

MSc in Artificial Intelligence Faculty of Engineering Uni-versity of LeedsMSc in Cognitiveamp Decision Sciences Psychology UniversityCollege LondonMSc in Cognitive Systems Language Learning and Reason-ing University of PotsdamMSc in Cognitive Science University of Osnabruck GermanyMSc in Cognitive PsychologyNeuropsychology School ofPsychology University of KentMSc in Logic Institute for Logic Language and ComputationUniversity of AmsterdamMSc in Mind Language amp Embodied Cognition School ofPhilosophy Psychology and Language Sciences University ofEdinburghMSc in Philosophy of Science Technology and Society Uni-versity of Twente The NetherlandsMRes in Cognitive Science and Humanities Language Com-munication and Organization Institute for Logic CognitionLanguage and Information University of the Basque Country(Donostia San Sebastian)OpenMind International School of Advanced Studies in Cog-nitive Sciences University of BucharestResearchMaster in Philosophy and Economics Erasmus Uni-versity Rotterdam The Netherlands

Jobs and Studentships

JobsAssistant Professor in Epistemology California State Uni-versity at Sacramento deadline until filledAssistant Professor in Logic amp Epistemology University ofNorth Carolina at Greensboro deadline until filledAssistant Professor in Philosophy of Medicine University ofNorth Carolina at Greensboro deadline until filledAssistant Professor in Philosophy of Science University ofFlorida deadline until filledLecturer in Actuarial Science University of California SantaBarbara deadline until filledProfessor in Philosophy of Science City University of NewYork deadline 3 DecemberResearch Associate in Bayesian modelling University ofSheffield deadline 12 DecemberResearch Fellow in Data Science University of Bristol dead-line 3 JanuaryLecturer in Statistical Science University College Londondeadline 9 February

100

  • Editorial
  • Features
  • News
  • Whats Hot in hellip
  • Events
  • Courses and Programmes
  • Jobs and Studentships

variant if relation symbols are added one has zero informationabout It fails to be language invariant if we add formal expres-sions for chances to the language It fails what we call ChanceInvariance Under this construal of invariant inferences it isthe Centre of Mass approach which simply picks out the cen-tre of mass of the acceptable probability functions which sat-isfies Chance Invariance Against all hope I think this peculiarfinding should get more press (2017 Invariant EquivocationJurgen Landes and George Masterton Erkenntnis 141-167)

Irsquove just completed my fourth postdoc On this projectwe aimed to assign a probability to the causal claim that adrug causes a particular adverse drug reaction While thereare many approaches marrying probabilities and causation weused the Bayesian network machinery to ldquomerelyrdquo assign aprobability to one causal hypothesis (2018 Epistemology ofCausal Inference in Pharmacology Jurgen Landes Barbara Os-imani Roland Poellinger European Journal for Philosophy ofScience 3-49) Irsquove been much interested in the confirma-tory value of varied evidence to the causal hypothesis in thisframework (2019 Variety of Evidence Jurgen Landes Erken-ntnis forthcoming DOI 101007s10670-018-0024-6 moremanuscripts under review)

HH Whatrsquos next thenJL Early this month Irsquove returned from paternity leave af-

ter the birth of our third child in August 2018 I now work onmy own research project on Evidence and Objective BayesianEpistemology In this three-year project I will investigate justi-fications and applications of amalgamating all the available ev-idence within the objective Bayesian approach Why and howdoes one (best) aggregate the available evidence

HH Thatrsquos a fascinating and timely question ndash we look for-ward to reading about your take on the problem To more per-sonal questions now You certainly crossed borders betweencountries disciplines and sectors over the past decade To-day the rise of rightwing populists and anti-EU nationalists arepushing forward a view which will make it increasingly harderfor the new generation of scientists to do that Any thoughts

JL Yes way too many thoughtsFirst we all donrsquot walk a mile in someone elsersquos shoes anymore before we judge In fact we donrsquot even seem to considera different opinion upbringing or socio-economic backgroundother than our own to be proper As a result societies are be-coming more and more partitioned into ever more divided sec-tions in which opinions from the outside no further penetrateThis facilitates populism around the world

Take us academics for example more and more of us (areforced to) work in different countries where we work at uni-versities in large cities and communicate with colleagues andincreasingly also staff in English all day every day Every-one has a decently waged also not necessarily stable workcontract We engage in academic cerebral exercises which areworlds apart from a typical day of early school leavers withoutany prospects for a decent job We all are living in our ownbubbles although we are ever more connected over the Inter-net

Irsquom deeply troubled by any nationalist movement indepen-dent of country and time We should always remember thatthere may be a time in the much too near future in which thebombs might be falling (the economy collapse natural disasterstrike etc) right where we stand now Wouldnrsquot it then be greatto have some friends who could shelter our kids To me thatrsquosa no-brainer

The best and clinching argument for the EU must be the longlong years of peace it has brought to Europe Just remembernot so long ago France and England fought a war that lasted for100 years Preventing wars for decades must be worth the tinybit of autonomy politicians (not countries) give up as well asall membership fees

As for crossing disciplines I can only advise to limit thenumber of cross-overs to a small number Every time you crossyou start from close to zero you lack a grasp of the interest-ing debates an understanding of the type of argument requiredto publish in top journals a personal network a well-stackedlibrary and so on Eventually you may well end up with a pub-lication record that does not impress in university departmentssince too many publications are outside their sphere of inter-est (2013 Embedding philosophers in the practices of sciencebringing humanities to the sciences Nancy Tuana Synthese1955-1973) We welcome back the above-mentioned bubbles

HH Can you see drawbacks in crossing bordersJL Crossing countries is a somewhat different matter That

is something I found deeply rewarding Furthermore it madevery clear to me that every country is populated by human be-ings which are not so different from each other Therersquos abso-lutely no reason for countries to get angry at another Itrsquos longoverdue that we all grew upHowever there are serious downsides to being an academic no-mad Irsquom not talking about the inevitable minor issues likecontributions to different pension systems hassle with differ-ent bureaucracies and trips to embassies Irsquom talking about theproblem of making friends and maintaining friendships over along period of time Just imagine every time you meet an in-teresting person you immediately think in a few yearrsquos timewe will live in different cities and probably different countriesThatrsquos not idealThe problem is worse for those of us who have children Rais-ing the young is extremely time consuming and tiring Nothaving a social support network in place (parentsgood friendsliving around the corner) because of a recent move does notmake things any easier Been there done it sent a postcard (3times)

HH Speaking of populisms we desperately need to getmore people interested in sound reasoning and argumentationYou said before itrsquos a shame it is not an innate interest of oursso can you think of actions we could take as a community toplay a leading role in urging the general public to pay attentionto reasoning

JL Most of our work is far far removed from the every-daylives of ordinary people I donrsquot think that we stand a seriouschance in engaging them in our academic brain activities

Grabbing the attention of inquisitive and excitable childrenwho may well recognise the value of good reasoning appearsmore effective in the long run to me Irsquove recently attended atalk by Andy Oxman He and his colleagues teach kids aboutrecognising and evaluating claims Scaling up this and othersuch initiatives is a way forward I believe in

While I think that getting the general public interested inreasoning as too big a task involving people facing compli-cated problems routinely at work makes more sense to me Thework by Vincenzo Crupi and others (eg 2017 Understandingand improving decisions in clinical medicine (I) Reasoningheuristics and error Internal and Emergency Medicine Vin-cenzo Crupi and Fabrizio Elia 689-691) is the type of project Iwould like to see more of

95

Finally engaging the interested public at a Ted Conferencesay is a good idea

News

Calls for PapersKnowing the Unknown Philosophical Perspectives on Igno-rance special issue of Synthese deadline 20 FebruaryHybrid Data and Knowledge Driven Decision Making underUncertainty special issue of Information Sciencesl deadline30 FebruaryThought Experiments in theHistory of Philosophy of Sciencespecial issue of HOPOS deadline 31 MarchFolk Psychology Pluralistic Approaches special issue ofSynthese deadline 15 MayImprecise Probabilities Logic and Rationality special issueof International Journal of Approximate Reasoning deadline 1June

Whatrsquos Hot in

Medieval ReasoningMany times students andcolleagues alike have askedme what would be a good in-troduction to medieval logicand I always struggle to finda good answer Usually Ibegin by telling them wherenot to start from Williamand Martha Kneale The De-velopment of Logic (Oxford1962) (And not becauseKneale amp Kneale is not aclassic groundbreaking textndash it is both Nor because it hasnrsquot aged gracefully ndash although ithasnrsquot aged particularly well at all But because one would getthe impression that medieval logicians ndash despite reaching stun-ning levels of apparent sophistication ndash got some very simplestuff very very wrong I donrsquot think that this is a fair evaluationnor is it a historically correct assessment Overall itrsquos better toleave Kneale amp Kneale on the side at the beginning perhaps tobe revisited later on) However pointing beginners in the direc-tion of a good introductory textbook is a trickier business Mostoverviews of medieval logic are either too partial or are fairlyobsolete ndash quite often both ndash and in many cases not really fitfor readers lacking a background in Ancient and Medieval phi-losophy Alexander Broadiersquos Introduction to Medieval Logic(2nd ed Oxford 1993) while being a fairly accessible and ba-sic text is a bit outdated and despite its title focuses heavily onthe first half of the 14th century Just as elementary (albeit justas partial and outdated as well) Paul Vincent Spadersquos ThoughtsWords and Things An Introduction to Late Medieval Logicand Semantic Theory (version 11 2002 httpspvspadecomLogicdocsthoughts1_1apdf) has both the faultsand virtues of an informal handbook put together from lecturenotes and circulated but never polished for publication JanPinborgrsquos Medieval Semantics Selected Studies on MedievalLogic and Grammar (English translation London 1984) while

not as basic has more thematic and chronologic breadth butunfortunately its age shows Normann Kretzmannrsquos AnthonyKennyrsquos and Jan Pinborgrsquos (eds) The Cambridge History ofLater Medieval Philosophy (Cambridge 1982) is largely de-voted to medieval logic and semantics to the point that itrsquos stillone of the most complete overviews around Some more re-cent overviews that collect a good number of rich and detailedarticles are Dov Gabbayrsquos and John Woodsrsquo (eds) Handbookof the History of Logic II Mediaeval and Renaissance Logic(Amsterdam 2008) and Catarina Dutilh Novaesrsquo and StephenReadrsquos (eds) Cambridge Companion to Medieval Logic (Cam-bridge 2016) Some other studies and anthologies come tomind ndash among those that are at least partially in English KlausJacobirsquos (ed) Argumentationstheorie (1993) Mikko Yrjonsu-urirsquos (ed) Medieval Formal Logic (2001) Dutilh Novaesrsquo For-malizing Medieval Logical Theories (2007) and Terry ParsonrsquosArticulating Medieval Logic (2014)

The above are all very interesting reads but ndash for the mostpart ndash they donrsquot seem to make for viable introductions It lookslike a Medieval Logic for Dummies (from 400 to 1450 ca) hasyet to be written ndash and until somebody does I wonrsquot be able togive a short simple answer to those who ask In the meantimetherersquos plenty of exciting stuff to skim through if you want todive into the subject headfirst

Graziana CiolaUCLA

Uncertain ReasoningTwenty years ago this De-cember James M Joycepublished ldquoA NonpragmaticVindication of Probabilismrdquo(Philosophy of Science1998) This paper inspireda whole new way to thinkabout epistemic norms Thebasic idea of rdquoscoring rulesrdquoor rdquoinaccuracy measuresrdquowas not new to Joyce in-deed the basics have beenaround since at least thefifties If yoursquore predicting whether or not it will rain and yourpredictions are given in the form of ldquoyesrdquo ldquonordquo answers itrsquosclear what counts as a good forecast if it rains you did wellif you said that it would and you did poorly if you said that itwouldnrsquot If instead you give your predictions in the form ofa probability of rain itrsquos less clear how to measure success Ifit rains and you said that the probability of rain was 90 thatis quite good ndash better than saying 80 worse than saying 95ndash but how exactly should one measure the quality of forecastsThis is important since we use probabilistic forecasts in ourdecision making and so it is important to know how to assessthe quality of a forecast In essence what one wants is someform of rdquodistance from the truthrdquo which one could use tomeasure how ldquocloserdquo your probabilistic predictions were Butthere are a great many different ways to measure the distancefrom the truth how can we choose between them

What Joyce pointed out was that while it might be unclearwhich measure of forecast quality is the one we ought to usethere are some features that they arguably ought to share andthat those common features are enough to build an argument

96

that our degrees of belief ought to obey the axioms of proba-bility theory (among other things) The details of Joycersquos orig-inal argument neednrsquot concern us here there have been manydifferent versions of the general approach since 1998 The ar-gumentative strategy ndash argue that certain principles of scoringrules constrain whatever epistemic utility function governs ourepistemic behaviour use these principles to prove what epis-temic norms we ought to obey ndash is one that has been takenup by many people and put to many uses Not just proba-bilism but conditionalisation and the principal principle havebeen justified in broadly this way This way of proceeding ar-gues Joyce is one that avoids the pitfalls and problems thatappear to plague the other standard way of justifying normsof credence namely ldquosure lossrdquo betting arguments (What aretypically called ldquoDutch book argumentsrdquo but for reasons I men-tioned last month Irsquom avoiding that term) For an up to dateoverview of the literature inspired by Joycersquos paper see RichardPettigrewrsquos recent book ldquoAccuracy and the Laws of Credencerdquo(OUP 2016)

This continues to be an interesting and fruitful area of re-search and Irsquoll return to the topic in a future column

Seamus BradleyPhilosophy University of Tilburg

Mathematical Philosophy

The current spectaculardevelopments in machinelearning go hand in handwith a growing popularinterest in the discipline Butinterest of formal philoso-phers in machine learningpredates the disciplinersquosrecent rise from an obscurebranch of artificial intelli-gence to its epitome activityThe reason is that machinelearning and formal epistemology share a concern with thesame fundamental question how to learn from data How toinfer conclusions from given data conclusions that go beyondthe data itself How could this be formalized or evenmdashthecharacteristic business of machine learningmdashautomatized

Thus several approaches within machine learning have beendirected by epistemologists at philosophical problems In someinstances like causal reasoning (Pearl Causality Cambridge2009) there has long been much interaction between philoso-phers and computer scientists As another example formallearning theory initiated by among others Putnam (J SymbLog 30(1) 49ndash57 1965) evolved into a technical field in com-puter science (Osherson et al Systems That Learn MIT 1985)and has more recently been developed further as a projectwithin philosophy of science (Kelly The Logic of ReliableInquiry Oxford 1996) Other instances are more unidirec-tional Schurz (Philos Sci 75 278ndash305 2008) has proposeda Reichenbachian justification of induction based on results incompetitive online learning (Cesa-Bianchi and Lugosi Predic-tion Learning and Games Cambridge 2006) Harman andKulkarni (Reliable Reasoning MIT 2006) have highlightedthe philosophical relevance of statistical learning theory theframework that underlies supervised learning (Vapnik Statisti-

cal Learning Theory Wiley 1998)On a more grand level a number of philosophers have argued

that machine learning prompts novel ways of doing philosophyof science This is in itself not new Thagard (MIT 1988) forinstance defended a computational philosophy of science in-spired by artificial intelligence Nor is this the norm Harmanand Kulkarni employ the statistical learning theory frameworkin a largely traditional analysis of the questions of inductionand its justification like the Goodman riddle and the role ofsimplicity But work within formal learning theory is explic-itly committed to a more pragmatic means-ends epistemology(Schulte Brit J Philos Sci 50(1) 1ndash31 1999) that insteadof the traditional focus on justifying inductive reasoning in gen-eral starts with the complexity of a given inductive problemthus determining the epistemic goals that are still feasible forit Wheeler (The Routledge Companion to Philosophy of So-cial Science 321ndash329 2017) goes much further still and de-clares ldquotraditional Enlightenment epistemologyrdquo bankrupt tobe replaced by the ldquoprimacy of practicerdquo of a pragmatist ma-chine epistemology

Certainly the concurrent advent of big data has contributedto the perception that machine learning is largely a practicalenterprise a matter of unleashing rough-and-ready algorithmson vast amounts of data In the most extreme version of thispicture theoretical analysis is not just hopelessly lagging be-hind it is superfluous and some have even speculated noto-riously about the consequences for scientific theorizing Thispresumed opposition between theory and practice however isalso a theme within machine learning Vapnik in the intro-ductory chapter to The Nature of Statistical Learning Theory(2nd edition Springer 2000) paints a picture of the history ofmachine learning where the practitioner is set against the theo-rist Vapnik chides the ldquoartificial intelligence hardlinersrdquo (ldquotheywho declared that lsquoComplex theories do not work simple al-gorithms dorsquordquo) for repeatedly making excessive promises andconcludes confidently that ldquoa new methodological situation inthe learning problem has developed where practical methodsare the result of a deep theoretical analysis of the statisticalbounds rather than the result of inventing new smart heuris-ticsrdquo Written still before the modern rehabilitation of neuralnetworks in the form of deep learning this now sounds overlyconfident The theme has indeed recently flared up again inthe community following Rahimi and Rechtrsquos speech at NIPS2017 in which they liken current machine learning research toalchemy

The predictable outrage notwithstanding Rahimi and Rechtnowhere commit to theory for the sake of theory they do nottake issue with the primacy of the pragmatic goal of findingalgorithms that work well in practice Their argument is thata theoretical understanding is important for this goal it is thestep back that facilitates so to speak a more efficient searchthrough the space of possible algorithms Continuous with thistheoretical work a further step back would be the foundationalwork that is the province of philosophers This concerns forinstance what it should mean when Rahimi and Recht speakabout ldquorigorous reliable verifiable knowledgerdquo and how thiscould be grounded in the relevant mathematics This also con-cerns the outer limits on learning algorithmsrsquo capabilities thesubject of my own work on universal prediction (University ofGroningen 2018)

The gap between theory and practice is particularly conspic-uous in what has been called a paradox of deep learning neural

97

networks perform astonishingly well in practice much betterthan they should in theory This is illustrated by Zhang et al(ldquoUnderstanding deep learning requires rethinking generaliza-tionrdquo ICLR 2017) with a simple experiment essentially theauthors took some standard architectures and data sets and thenchanged the data by reshuffling the labels in a random mannerThis guaranteed that there was no relation between instancesand labels yet the networks still managed to attain extremelylow training error entailing a wide divergence between train-ing error and generalization error (the latter must still be veryhigh the labeling is fully random so there is nothing to learn)But explanations based on statistical learning theory rely on aconnection between training error (or more generally the archi-tecture of the network) and generalization error wherefore thedemonstrated possibility of varying the latter while keeping theformer constant implies that these explanations simply do notapply Depending on onersquos perspective one could take this asanother illustration of the impotence of theory or of the direneed for a greater theoretical effort In either case the properway to understand deep learning is an important ongoing de-bate in the machine learning community that merits attentionfrom philosophers too

Tom SterkenburgMunich Centre for Mathematical Philosophy

Evidence-Based Medicine

Precision medicine is not new but it is currently a hot topic inEBM Often used interchangeably with personalised medicinethis new approach to healthcare aims to individualise treat-ments by recognizing the ways in which patients can differ per-son to person in the nature of disease and in patient responseto treatments It is well known that patients will respond dif-ferently to treatments that are effective for other people withthe same disease It is being increasingly recognised that thisis due to biological differences in the patients and the diseasesthemselves often at the genetic level Two very recent papersarticulate what precision medicine aims to do identify the ge-netic differences among patients with a certain disease and usenew technologies to enable us to make sense of the level ofcomplexity associated with using genetics to make predictions

One study classified subtypes of acute myeloid leukemia(AML) by their gene expression and DNA-methylation pat-terns AML is a single disease as it has a specific characteri-zation ldquoblocked myeloid lineage differentiation and accumu-lation of leukemic blast cellsrdquo But it is a ldquohighly heterogeneousdiseaserdquo as it can be caused by many different types of geneticalterations These different genetic alterations define separatesub-types of AML and patients with different sub-types willrequire different specific treatments This study identified pat-terns in the expression of genes in tumor cells from AML pa-tients which revealed gene regulatory networks that are spe-cific to the different sub-types of AML There were a numberof recurrent alterations important for lukemic growth identifiedthat present possibilities for targeted treatment Without thisknowledge patients may receive treatment that will not be ef-fective as it is not targeted against their specific AML subtype

The results of this study are claimed to be able to base tar-geted treatments for specific AML sub-types It is importantto note here that while precision medicine aims to take resultssuch as these and use them to target individual patients it is

actually sub-populations of patients with specific AML sub-types that can be targeted rather than individual patients per seDrugs will not be tailored to any one individual but novel drugscan be designed with specific sub-populations in mind and ex-tant drugs can be better targeted at specific sub-populationsThis is why precision medicine may be a better term than per-sonalised medicine as these techniques allow us to be moreprecise in targeting treatments to any one individual This mayseem like a small semantic difference but being clear about thistempers the ability to make grandiose claims about treating in-dividuals - it still might be the case that precision treatmentsare not effective for individuals due to chance or other factorsboth biological and environmental

Difficulties face the implementation of this research bothbroadly and specifically A difficulty facing precision medicineis practical in nature the level of complexity associated withusing results from disciplines such as genetics One way inwhich precision medicine hopes to overcome this challenge isthrough the use of technology in particular AI A practical dif-ficulty specific to oncology is that ldquotumors change over timeprogressing from benign to malignant becoming metastaticand developing resistance to certain therapiesrdquo Essentiallythe tumor cells evolve resulting in lsquointratumour heterogeneityrsquoThis means that due to tumour evolution using genetic mark-ers of cancer subtypes may quickly not be useful for targetingtherapies An enterprise hoping to provide solutions to bothdifficulties is the REVOLVER system that has recently pub-lished results on predicting changes in tumour evolution usingan AI system This system uses a machine-learning approachknown as transfer learning (TL)19 to predict the next steps inspecific tumor cell evolution Tumor evolution seems to be pre-dictable due to the prognostic value of histopathological stag-ing and molecular markers Problems facing using this data topredict tumor evolution are the complexity of patterns in theavailable data and stochastic forces in the mechanisms of tu-mor cell evolution The REVOLVER system can deal betterwith complexity and stochasticity than a human reasoner Thisenables more precise information for specific cancer patients asto what might happen during the course of their disease

I have written about AI in this column before and with thatcase (triaging in opthamology) the system was again used towork with large quantities of data to improve the efficiency ofhealth care delivery Precision medicine offers much in the wayof improving how we treat severe diseases but as the data setsget larger and more complex we will increasingly have to turnto AI It is promising that there is development in this fieldthat can meet this challenge Interestingly at present the datasets are not actually large enough for the AI systems to providethe high degree of precision we require But with precisionmedicine research considered a major growth area in medicinewe will likely not have to wait too long to overcome this prob-lem

DJ Auker-HowlettPhilosophy Kent

98

Events

December

KEiMS Conference on Knowledge Exchange in the Mathe-matical Sciences Aston University 3ndash4 DecemberCvC Causation vs Constitution Loosening the FrictionBergen 3ndash4 DecemberAxiMet Axiomatizing MetatheorymdashTruth Provability andBeyond Salzburg Austria 6ndash7 DecemberML4H Workshop on Machine Learning for Health MontrealCanada 8 DecemberDiI Workshop on Disagreement in Inquiry University of Tartu8 DecemberRLPO Reinforcement Learning Under Partial ObservabilityMontreal Canada 8 DecemberRPiB Real Patterns University of Barcelona 14 DecemberWrsquosBaD Whatrsquos so Bad About Dialetheism Kyoto Univer-sity Japan 15ndash17 December

January

PoMaL Graduate Conference on the Philosophy of Mathemat-ics and Logic University of Cambridge 19ndash20 JanuaryFiSS Foundations in Social SciencemdashMechanisms ActionsFunctions University of Duisburg-Essen Germany 24ndash25 Jan-uaryVoUiS Varieties of Understanding in Science Utrecht TheNetherlands 24ndash25 January

February

DMaMG Dark Matter and Modified Gravity ConferenceAachen Germany 6ndash8 February

Courses and Programmes

Courses

SSA Summer School on Argumentation Computational andLinguistic Perspectives on Argumentation Warsaw Poland 6ndash10 September

Programmes

APhil MAPhD in Analytic Philosophy University ofBarcelonaMaster Programme MA in Pure and Applied Logic Univer-sity of BarcelonaDoctoral Programme in Philosophy Language Mind andPractice Department of Philosophy University of ZurichSwitzerlandDoctoral Programme in Philosophy Department of Philoso-phy University of Milan ItalyLogiCS Joint doctoral program on Logical Methods in Com-puter Science TU Wien TU Graz and JKU Linz AustriaHPSM MA in the History and Philosophy of Science andMedicine Durham UniversityMaster Programme in Statistics University College DublinLoPhiSC Master in Logic Philosophy of Science and Epis-temology Pantheon-Sorbonne University (Paris 1) and Paris-Sorbonne University (Paris 4)Master Programme in Artificial Intelligence Radboud Uni-versity Nijmegen the NetherlandsMaster Programme Philosophy and Economics Institute ofPhilosophy University of BayreuthMA in Cognitive Science School of Politics InternationalStudies and Philosophy Queenrsquos University BelfastMA in Logic and the Philosophy ofMathematics Departmentof Philosophy University of BristolMA Programmes in Philosophy of Science University ofLeedsMA in Logic and Philosophy of Science Faculty of PhilosophyPhilosophy of Science and Study of Religion LMU MunichMA in Logic and Theory of Science Department of Logic ofthe Eotvos Lorand University Budapest HungaryMA in Metaphysics Language and Mind Department of Phi-losophy University of LiverpoolMA inMind Brain and Learning Westminster Institute of Ed-ucation Oxford Brookes UniversityMA in Philosophy by research Tilburg UniversityMA in Philosophy Science and Society TiLPS Tilburg Uni-versityMA in Philosophy of Biological and Cognitive Sciences De-partment of Philosophy University of BristolMA in Rhetoric School of Journalism Media and Communi-cation University of Central LancashireMA programmes in Philosophy of Language and Linguisticsand Philosophy of Mind and Psychology University of Birm-inghamMRes in Methods and Practices of Philosophical ResearchNorthern Institute of Philosophy University of AberdeenMSc in Applied Statistics Department of Economics Mathe-matics and Statistics Birkbeck University of LondonMSc in Applied Statistics and Datamining School of Mathe-matics and Statistics University of St Andrews

99

MSc in Artificial Intelligence Faculty of Engineering Uni-versity of LeedsMSc in Cognitiveamp Decision Sciences Psychology UniversityCollege LondonMSc in Cognitive Systems Language Learning and Reason-ing University of PotsdamMSc in Cognitive Science University of Osnabruck GermanyMSc in Cognitive PsychologyNeuropsychology School ofPsychology University of KentMSc in Logic Institute for Logic Language and ComputationUniversity of AmsterdamMSc in Mind Language amp Embodied Cognition School ofPhilosophy Psychology and Language Sciences University ofEdinburghMSc in Philosophy of Science Technology and Society Uni-versity of Twente The NetherlandsMRes in Cognitive Science and Humanities Language Com-munication and Organization Institute for Logic CognitionLanguage and Information University of the Basque Country(Donostia San Sebastian)OpenMind International School of Advanced Studies in Cog-nitive Sciences University of BucharestResearchMaster in Philosophy and Economics Erasmus Uni-versity Rotterdam The Netherlands

Jobs and Studentships

JobsAssistant Professor in Epistemology California State Uni-versity at Sacramento deadline until filledAssistant Professor in Logic amp Epistemology University ofNorth Carolina at Greensboro deadline until filledAssistant Professor in Philosophy of Medicine University ofNorth Carolina at Greensboro deadline until filledAssistant Professor in Philosophy of Science University ofFlorida deadline until filledLecturer in Actuarial Science University of California SantaBarbara deadline until filledProfessor in Philosophy of Science City University of NewYork deadline 3 DecemberResearch Associate in Bayesian modelling University ofSheffield deadline 12 DecemberResearch Fellow in Data Science University of Bristol dead-line 3 JanuaryLecturer in Statistical Science University College Londondeadline 9 February

100

  • Editorial
  • Features
  • News
  • Whats Hot in hellip
  • Events
  • Courses and Programmes
  • Jobs and Studentships

Finally engaging the interested public at a Ted Conferencesay is a good idea

News

Calls for PapersKnowing the Unknown Philosophical Perspectives on Igno-rance special issue of Synthese deadline 20 FebruaryHybrid Data and Knowledge Driven Decision Making underUncertainty special issue of Information Sciencesl deadline30 FebruaryThought Experiments in theHistory of Philosophy of Sciencespecial issue of HOPOS deadline 31 MarchFolk Psychology Pluralistic Approaches special issue ofSynthese deadline 15 MayImprecise Probabilities Logic and Rationality special issueof International Journal of Approximate Reasoning deadline 1June

Whatrsquos Hot in

Medieval ReasoningMany times students andcolleagues alike have askedme what would be a good in-troduction to medieval logicand I always struggle to finda good answer Usually Ibegin by telling them wherenot to start from Williamand Martha Kneale The De-velopment of Logic (Oxford1962) (And not becauseKneale amp Kneale is not aclassic groundbreaking textndash it is both Nor because it hasnrsquot aged gracefully ndash although ithasnrsquot aged particularly well at all But because one would getthe impression that medieval logicians ndash despite reaching stun-ning levels of apparent sophistication ndash got some very simplestuff very very wrong I donrsquot think that this is a fair evaluationnor is it a historically correct assessment Overall itrsquos better toleave Kneale amp Kneale on the side at the beginning perhaps tobe revisited later on) However pointing beginners in the direc-tion of a good introductory textbook is a trickier business Mostoverviews of medieval logic are either too partial or are fairlyobsolete ndash quite often both ndash and in many cases not really fitfor readers lacking a background in Ancient and Medieval phi-losophy Alexander Broadiersquos Introduction to Medieval Logic(2nd ed Oxford 1993) while being a fairly accessible and ba-sic text is a bit outdated and despite its title focuses heavily onthe first half of the 14th century Just as elementary (albeit justas partial and outdated as well) Paul Vincent Spadersquos ThoughtsWords and Things An Introduction to Late Medieval Logicand Semantic Theory (version 11 2002 httpspvspadecomLogicdocsthoughts1_1apdf) has both the faultsand virtues of an informal handbook put together from lecturenotes and circulated but never polished for publication JanPinborgrsquos Medieval Semantics Selected Studies on MedievalLogic and Grammar (English translation London 1984) while

not as basic has more thematic and chronologic breadth butunfortunately its age shows Normann Kretzmannrsquos AnthonyKennyrsquos and Jan Pinborgrsquos (eds) The Cambridge History ofLater Medieval Philosophy (Cambridge 1982) is largely de-voted to medieval logic and semantics to the point that itrsquos stillone of the most complete overviews around Some more re-cent overviews that collect a good number of rich and detailedarticles are Dov Gabbayrsquos and John Woodsrsquo (eds) Handbookof the History of Logic II Mediaeval and Renaissance Logic(Amsterdam 2008) and Catarina Dutilh Novaesrsquo and StephenReadrsquos (eds) Cambridge Companion to Medieval Logic (Cam-bridge 2016) Some other studies and anthologies come tomind ndash among those that are at least partially in English KlausJacobirsquos (ed) Argumentationstheorie (1993) Mikko Yrjonsu-urirsquos (ed) Medieval Formal Logic (2001) Dutilh Novaesrsquo For-malizing Medieval Logical Theories (2007) and Terry ParsonrsquosArticulating Medieval Logic (2014)

The above are all very interesting reads but ndash for the mostpart ndash they donrsquot seem to make for viable introductions It lookslike a Medieval Logic for Dummies (from 400 to 1450 ca) hasyet to be written ndash and until somebody does I wonrsquot be able togive a short simple answer to those who ask In the meantimetherersquos plenty of exciting stuff to skim through if you want todive into the subject headfirst

Graziana CiolaUCLA

Uncertain ReasoningTwenty years ago this De-cember James M Joycepublished ldquoA NonpragmaticVindication of Probabilismrdquo(Philosophy of Science1998) This paper inspireda whole new way to thinkabout epistemic norms Thebasic idea of rdquoscoring rulesrdquoor rdquoinaccuracy measuresrdquowas not new to Joyce in-deed the basics have beenaround since at least thefifties If yoursquore predicting whether or not it will rain and yourpredictions are given in the form of ldquoyesrdquo ldquonordquo answers itrsquosclear what counts as a good forecast if it rains you did wellif you said that it would and you did poorly if you said that itwouldnrsquot If instead you give your predictions in the form ofa probability of rain itrsquos less clear how to measure success Ifit rains and you said that the probability of rain was 90 thatis quite good ndash better than saying 80 worse than saying 95ndash but how exactly should one measure the quality of forecastsThis is important since we use probabilistic forecasts in ourdecision making and so it is important to know how to assessthe quality of a forecast In essence what one wants is someform of rdquodistance from the truthrdquo which one could use tomeasure how ldquocloserdquo your probabilistic predictions were Butthere are a great many different ways to measure the distancefrom the truth how can we choose between them

What Joyce pointed out was that while it might be unclearwhich measure of forecast quality is the one we ought to usethere are some features that they arguably ought to share andthat those common features are enough to build an argument

96

that our degrees of belief ought to obey the axioms of proba-bility theory (among other things) The details of Joycersquos orig-inal argument neednrsquot concern us here there have been manydifferent versions of the general approach since 1998 The ar-gumentative strategy ndash argue that certain principles of scoringrules constrain whatever epistemic utility function governs ourepistemic behaviour use these principles to prove what epis-temic norms we ought to obey ndash is one that has been takenup by many people and put to many uses Not just proba-bilism but conditionalisation and the principal principle havebeen justified in broadly this way This way of proceeding ar-gues Joyce is one that avoids the pitfalls and problems thatappear to plague the other standard way of justifying normsof credence namely ldquosure lossrdquo betting arguments (What aretypically called ldquoDutch book argumentsrdquo but for reasons I men-tioned last month Irsquom avoiding that term) For an up to dateoverview of the literature inspired by Joycersquos paper see RichardPettigrewrsquos recent book ldquoAccuracy and the Laws of Credencerdquo(OUP 2016)

This continues to be an interesting and fruitful area of re-search and Irsquoll return to the topic in a future column

Seamus BradleyPhilosophy University of Tilburg

Mathematical Philosophy

The current spectaculardevelopments in machinelearning go hand in handwith a growing popularinterest in the discipline Butinterest of formal philoso-phers in machine learningpredates the disciplinersquosrecent rise from an obscurebranch of artificial intelli-gence to its epitome activityThe reason is that machinelearning and formal epistemology share a concern with thesame fundamental question how to learn from data How toinfer conclusions from given data conclusions that go beyondthe data itself How could this be formalized or evenmdashthecharacteristic business of machine learningmdashautomatized

Thus several approaches within machine learning have beendirected by epistemologists at philosophical problems In someinstances like causal reasoning (Pearl Causality Cambridge2009) there has long been much interaction between philoso-phers and computer scientists As another example formallearning theory initiated by among others Putnam (J SymbLog 30(1) 49ndash57 1965) evolved into a technical field in com-puter science (Osherson et al Systems That Learn MIT 1985)and has more recently been developed further as a projectwithin philosophy of science (Kelly The Logic of ReliableInquiry Oxford 1996) Other instances are more unidirec-tional Schurz (Philos Sci 75 278ndash305 2008) has proposeda Reichenbachian justification of induction based on results incompetitive online learning (Cesa-Bianchi and Lugosi Predic-tion Learning and Games Cambridge 2006) Harman andKulkarni (Reliable Reasoning MIT 2006) have highlightedthe philosophical relevance of statistical learning theory theframework that underlies supervised learning (Vapnik Statisti-

cal Learning Theory Wiley 1998)On a more grand level a number of philosophers have argued

that machine learning prompts novel ways of doing philosophyof science This is in itself not new Thagard (MIT 1988) forinstance defended a computational philosophy of science in-spired by artificial intelligence Nor is this the norm Harmanand Kulkarni employ the statistical learning theory frameworkin a largely traditional analysis of the questions of inductionand its justification like the Goodman riddle and the role ofsimplicity But work within formal learning theory is explic-itly committed to a more pragmatic means-ends epistemology(Schulte Brit J Philos Sci 50(1) 1ndash31 1999) that insteadof the traditional focus on justifying inductive reasoning in gen-eral starts with the complexity of a given inductive problemthus determining the epistemic goals that are still feasible forit Wheeler (The Routledge Companion to Philosophy of So-cial Science 321ndash329 2017) goes much further still and de-clares ldquotraditional Enlightenment epistemologyrdquo bankrupt tobe replaced by the ldquoprimacy of practicerdquo of a pragmatist ma-chine epistemology

Certainly the concurrent advent of big data has contributedto the perception that machine learning is largely a practicalenterprise a matter of unleashing rough-and-ready algorithmson vast amounts of data In the most extreme version of thispicture theoretical analysis is not just hopelessly lagging be-hind it is superfluous and some have even speculated noto-riously about the consequences for scientific theorizing Thispresumed opposition between theory and practice however isalso a theme within machine learning Vapnik in the intro-ductory chapter to The Nature of Statistical Learning Theory(2nd edition Springer 2000) paints a picture of the history ofmachine learning where the practitioner is set against the theo-rist Vapnik chides the ldquoartificial intelligence hardlinersrdquo (ldquotheywho declared that lsquoComplex theories do not work simple al-gorithms dorsquordquo) for repeatedly making excessive promises andconcludes confidently that ldquoa new methodological situation inthe learning problem has developed where practical methodsare the result of a deep theoretical analysis of the statisticalbounds rather than the result of inventing new smart heuris-ticsrdquo Written still before the modern rehabilitation of neuralnetworks in the form of deep learning this now sounds overlyconfident The theme has indeed recently flared up again inthe community following Rahimi and Rechtrsquos speech at NIPS2017 in which they liken current machine learning research toalchemy

The predictable outrage notwithstanding Rahimi and Rechtnowhere commit to theory for the sake of theory they do nottake issue with the primacy of the pragmatic goal of findingalgorithms that work well in practice Their argument is thata theoretical understanding is important for this goal it is thestep back that facilitates so to speak a more efficient searchthrough the space of possible algorithms Continuous with thistheoretical work a further step back would be the foundationalwork that is the province of philosophers This concerns forinstance what it should mean when Rahimi and Recht speakabout ldquorigorous reliable verifiable knowledgerdquo and how thiscould be grounded in the relevant mathematics This also con-cerns the outer limits on learning algorithmsrsquo capabilities thesubject of my own work on universal prediction (University ofGroningen 2018)

The gap between theory and practice is particularly conspic-uous in what has been called a paradox of deep learning neural

97

networks perform astonishingly well in practice much betterthan they should in theory This is illustrated by Zhang et al(ldquoUnderstanding deep learning requires rethinking generaliza-tionrdquo ICLR 2017) with a simple experiment essentially theauthors took some standard architectures and data sets and thenchanged the data by reshuffling the labels in a random mannerThis guaranteed that there was no relation between instancesand labels yet the networks still managed to attain extremelylow training error entailing a wide divergence between train-ing error and generalization error (the latter must still be veryhigh the labeling is fully random so there is nothing to learn)But explanations based on statistical learning theory rely on aconnection between training error (or more generally the archi-tecture of the network) and generalization error wherefore thedemonstrated possibility of varying the latter while keeping theformer constant implies that these explanations simply do notapply Depending on onersquos perspective one could take this asanother illustration of the impotence of theory or of the direneed for a greater theoretical effort In either case the properway to understand deep learning is an important ongoing de-bate in the machine learning community that merits attentionfrom philosophers too

Tom SterkenburgMunich Centre for Mathematical Philosophy

Evidence-Based Medicine

Precision medicine is not new but it is currently a hot topic inEBM Often used interchangeably with personalised medicinethis new approach to healthcare aims to individualise treat-ments by recognizing the ways in which patients can differ per-son to person in the nature of disease and in patient responseto treatments It is well known that patients will respond dif-ferently to treatments that are effective for other people withthe same disease It is being increasingly recognised that thisis due to biological differences in the patients and the diseasesthemselves often at the genetic level Two very recent papersarticulate what precision medicine aims to do identify the ge-netic differences among patients with a certain disease and usenew technologies to enable us to make sense of the level ofcomplexity associated with using genetics to make predictions

One study classified subtypes of acute myeloid leukemia(AML) by their gene expression and DNA-methylation pat-terns AML is a single disease as it has a specific characteri-zation ldquoblocked myeloid lineage differentiation and accumu-lation of leukemic blast cellsrdquo But it is a ldquohighly heterogeneousdiseaserdquo as it can be caused by many different types of geneticalterations These different genetic alterations define separatesub-types of AML and patients with different sub-types willrequire different specific treatments This study identified pat-terns in the expression of genes in tumor cells from AML pa-tients which revealed gene regulatory networks that are spe-cific to the different sub-types of AML There were a numberof recurrent alterations important for lukemic growth identifiedthat present possibilities for targeted treatment Without thisknowledge patients may receive treatment that will not be ef-fective as it is not targeted against their specific AML subtype

The results of this study are claimed to be able to base tar-geted treatments for specific AML sub-types It is importantto note here that while precision medicine aims to take resultssuch as these and use them to target individual patients it is

actually sub-populations of patients with specific AML sub-types that can be targeted rather than individual patients per seDrugs will not be tailored to any one individual but novel drugscan be designed with specific sub-populations in mind and ex-tant drugs can be better targeted at specific sub-populationsThis is why precision medicine may be a better term than per-sonalised medicine as these techniques allow us to be moreprecise in targeting treatments to any one individual This mayseem like a small semantic difference but being clear about thistempers the ability to make grandiose claims about treating in-dividuals - it still might be the case that precision treatmentsare not effective for individuals due to chance or other factorsboth biological and environmental

Difficulties face the implementation of this research bothbroadly and specifically A difficulty facing precision medicineis practical in nature the level of complexity associated withusing results from disciplines such as genetics One way inwhich precision medicine hopes to overcome this challenge isthrough the use of technology in particular AI A practical dif-ficulty specific to oncology is that ldquotumors change over timeprogressing from benign to malignant becoming metastaticand developing resistance to certain therapiesrdquo Essentiallythe tumor cells evolve resulting in lsquointratumour heterogeneityrsquoThis means that due to tumour evolution using genetic mark-ers of cancer subtypes may quickly not be useful for targetingtherapies An enterprise hoping to provide solutions to bothdifficulties is the REVOLVER system that has recently pub-lished results on predicting changes in tumour evolution usingan AI system This system uses a machine-learning approachknown as transfer learning (TL)19 to predict the next steps inspecific tumor cell evolution Tumor evolution seems to be pre-dictable due to the prognostic value of histopathological stag-ing and molecular markers Problems facing using this data topredict tumor evolution are the complexity of patterns in theavailable data and stochastic forces in the mechanisms of tu-mor cell evolution The REVOLVER system can deal betterwith complexity and stochasticity than a human reasoner Thisenables more precise information for specific cancer patients asto what might happen during the course of their disease

I have written about AI in this column before and with thatcase (triaging in opthamology) the system was again used towork with large quantities of data to improve the efficiency ofhealth care delivery Precision medicine offers much in the wayof improving how we treat severe diseases but as the data setsget larger and more complex we will increasingly have to turnto AI It is promising that there is development in this fieldthat can meet this challenge Interestingly at present the datasets are not actually large enough for the AI systems to providethe high degree of precision we require But with precisionmedicine research considered a major growth area in medicinewe will likely not have to wait too long to overcome this prob-lem

DJ Auker-HowlettPhilosophy Kent

98

Events

December

KEiMS Conference on Knowledge Exchange in the Mathe-matical Sciences Aston University 3ndash4 DecemberCvC Causation vs Constitution Loosening the FrictionBergen 3ndash4 DecemberAxiMet Axiomatizing MetatheorymdashTruth Provability andBeyond Salzburg Austria 6ndash7 DecemberML4H Workshop on Machine Learning for Health MontrealCanada 8 DecemberDiI Workshop on Disagreement in Inquiry University of Tartu8 DecemberRLPO Reinforcement Learning Under Partial ObservabilityMontreal Canada 8 DecemberRPiB Real Patterns University of Barcelona 14 DecemberWrsquosBaD Whatrsquos so Bad About Dialetheism Kyoto Univer-sity Japan 15ndash17 December

January

PoMaL Graduate Conference on the Philosophy of Mathemat-ics and Logic University of Cambridge 19ndash20 JanuaryFiSS Foundations in Social SciencemdashMechanisms ActionsFunctions University of Duisburg-Essen Germany 24ndash25 Jan-uaryVoUiS Varieties of Understanding in Science Utrecht TheNetherlands 24ndash25 January

February

DMaMG Dark Matter and Modified Gravity ConferenceAachen Germany 6ndash8 February

Courses and Programmes

Courses

SSA Summer School on Argumentation Computational andLinguistic Perspectives on Argumentation Warsaw Poland 6ndash10 September

Programmes

APhil MAPhD in Analytic Philosophy University ofBarcelonaMaster Programme MA in Pure and Applied Logic Univer-sity of BarcelonaDoctoral Programme in Philosophy Language Mind andPractice Department of Philosophy University of ZurichSwitzerlandDoctoral Programme in Philosophy Department of Philoso-phy University of Milan ItalyLogiCS Joint doctoral program on Logical Methods in Com-puter Science TU Wien TU Graz and JKU Linz AustriaHPSM MA in the History and Philosophy of Science andMedicine Durham UniversityMaster Programme in Statistics University College DublinLoPhiSC Master in Logic Philosophy of Science and Epis-temology Pantheon-Sorbonne University (Paris 1) and Paris-Sorbonne University (Paris 4)Master Programme in Artificial Intelligence Radboud Uni-versity Nijmegen the NetherlandsMaster Programme Philosophy and Economics Institute ofPhilosophy University of BayreuthMA in Cognitive Science School of Politics InternationalStudies and Philosophy Queenrsquos University BelfastMA in Logic and the Philosophy ofMathematics Departmentof Philosophy University of BristolMA Programmes in Philosophy of Science University ofLeedsMA in Logic and Philosophy of Science Faculty of PhilosophyPhilosophy of Science and Study of Religion LMU MunichMA in Logic and Theory of Science Department of Logic ofthe Eotvos Lorand University Budapest HungaryMA in Metaphysics Language and Mind Department of Phi-losophy University of LiverpoolMA inMind Brain and Learning Westminster Institute of Ed-ucation Oxford Brookes UniversityMA in Philosophy by research Tilburg UniversityMA in Philosophy Science and Society TiLPS Tilburg Uni-versityMA in Philosophy of Biological and Cognitive Sciences De-partment of Philosophy University of BristolMA in Rhetoric School of Journalism Media and Communi-cation University of Central LancashireMA programmes in Philosophy of Language and Linguisticsand Philosophy of Mind and Psychology University of Birm-inghamMRes in Methods and Practices of Philosophical ResearchNorthern Institute of Philosophy University of AberdeenMSc in Applied Statistics Department of Economics Mathe-matics and Statistics Birkbeck University of LondonMSc in Applied Statistics and Datamining School of Mathe-matics and Statistics University of St Andrews

99

MSc in Artificial Intelligence Faculty of Engineering Uni-versity of LeedsMSc in Cognitiveamp Decision Sciences Psychology UniversityCollege LondonMSc in Cognitive Systems Language Learning and Reason-ing University of PotsdamMSc in Cognitive Science University of Osnabruck GermanyMSc in Cognitive PsychologyNeuropsychology School ofPsychology University of KentMSc in Logic Institute for Logic Language and ComputationUniversity of AmsterdamMSc in Mind Language amp Embodied Cognition School ofPhilosophy Psychology and Language Sciences University ofEdinburghMSc in Philosophy of Science Technology and Society Uni-versity of Twente The NetherlandsMRes in Cognitive Science and Humanities Language Com-munication and Organization Institute for Logic CognitionLanguage and Information University of the Basque Country(Donostia San Sebastian)OpenMind International School of Advanced Studies in Cog-nitive Sciences University of BucharestResearchMaster in Philosophy and Economics Erasmus Uni-versity Rotterdam The Netherlands

Jobs and Studentships

JobsAssistant Professor in Epistemology California State Uni-versity at Sacramento deadline until filledAssistant Professor in Logic amp Epistemology University ofNorth Carolina at Greensboro deadline until filledAssistant Professor in Philosophy of Medicine University ofNorth Carolina at Greensboro deadline until filledAssistant Professor in Philosophy of Science University ofFlorida deadline until filledLecturer in Actuarial Science University of California SantaBarbara deadline until filledProfessor in Philosophy of Science City University of NewYork deadline 3 DecemberResearch Associate in Bayesian modelling University ofSheffield deadline 12 DecemberResearch Fellow in Data Science University of Bristol dead-line 3 JanuaryLecturer in Statistical Science University College Londondeadline 9 February

100

  • Editorial
  • Features
  • News
  • Whats Hot in hellip
  • Events
  • Courses and Programmes
  • Jobs and Studentships

that our degrees of belief ought to obey the axioms of proba-bility theory (among other things) The details of Joycersquos orig-inal argument neednrsquot concern us here there have been manydifferent versions of the general approach since 1998 The ar-gumentative strategy ndash argue that certain principles of scoringrules constrain whatever epistemic utility function governs ourepistemic behaviour use these principles to prove what epis-temic norms we ought to obey ndash is one that has been takenup by many people and put to many uses Not just proba-bilism but conditionalisation and the principal principle havebeen justified in broadly this way This way of proceeding ar-gues Joyce is one that avoids the pitfalls and problems thatappear to plague the other standard way of justifying normsof credence namely ldquosure lossrdquo betting arguments (What aretypically called ldquoDutch book argumentsrdquo but for reasons I men-tioned last month Irsquom avoiding that term) For an up to dateoverview of the literature inspired by Joycersquos paper see RichardPettigrewrsquos recent book ldquoAccuracy and the Laws of Credencerdquo(OUP 2016)

This continues to be an interesting and fruitful area of re-search and Irsquoll return to the topic in a future column

Seamus BradleyPhilosophy University of Tilburg

Mathematical Philosophy

The current spectaculardevelopments in machinelearning go hand in handwith a growing popularinterest in the discipline Butinterest of formal philoso-phers in machine learningpredates the disciplinersquosrecent rise from an obscurebranch of artificial intelli-gence to its epitome activityThe reason is that machinelearning and formal epistemology share a concern with thesame fundamental question how to learn from data How toinfer conclusions from given data conclusions that go beyondthe data itself How could this be formalized or evenmdashthecharacteristic business of machine learningmdashautomatized

Thus several approaches within machine learning have beendirected by epistemologists at philosophical problems In someinstances like causal reasoning (Pearl Causality Cambridge2009) there has long been much interaction between philoso-phers and computer scientists As another example formallearning theory initiated by among others Putnam (J SymbLog 30(1) 49ndash57 1965) evolved into a technical field in com-puter science (Osherson et al Systems That Learn MIT 1985)and has more recently been developed further as a projectwithin philosophy of science (Kelly The Logic of ReliableInquiry Oxford 1996) Other instances are more unidirec-tional Schurz (Philos Sci 75 278ndash305 2008) has proposeda Reichenbachian justification of induction based on results incompetitive online learning (Cesa-Bianchi and Lugosi Predic-tion Learning and Games Cambridge 2006) Harman andKulkarni (Reliable Reasoning MIT 2006) have highlightedthe philosophical relevance of statistical learning theory theframework that underlies supervised learning (Vapnik Statisti-

cal Learning Theory Wiley 1998)On a more grand level a number of philosophers have argued

that machine learning prompts novel ways of doing philosophyof science This is in itself not new Thagard (MIT 1988) forinstance defended a computational philosophy of science in-spired by artificial intelligence Nor is this the norm Harmanand Kulkarni employ the statistical learning theory frameworkin a largely traditional analysis of the questions of inductionand its justification like the Goodman riddle and the role ofsimplicity But work within formal learning theory is explic-itly committed to a more pragmatic means-ends epistemology(Schulte Brit J Philos Sci 50(1) 1ndash31 1999) that insteadof the traditional focus on justifying inductive reasoning in gen-eral starts with the complexity of a given inductive problemthus determining the epistemic goals that are still feasible forit Wheeler (The Routledge Companion to Philosophy of So-cial Science 321ndash329 2017) goes much further still and de-clares ldquotraditional Enlightenment epistemologyrdquo bankrupt tobe replaced by the ldquoprimacy of practicerdquo of a pragmatist ma-chine epistemology

Certainly the concurrent advent of big data has contributedto the perception that machine learning is largely a practicalenterprise a matter of unleashing rough-and-ready algorithmson vast amounts of data In the most extreme version of thispicture theoretical analysis is not just hopelessly lagging be-hind it is superfluous and some have even speculated noto-riously about the consequences for scientific theorizing Thispresumed opposition between theory and practice however isalso a theme within machine learning Vapnik in the intro-ductory chapter to The Nature of Statistical Learning Theory(2nd edition Springer 2000) paints a picture of the history ofmachine learning where the practitioner is set against the theo-rist Vapnik chides the ldquoartificial intelligence hardlinersrdquo (ldquotheywho declared that lsquoComplex theories do not work simple al-gorithms dorsquordquo) for repeatedly making excessive promises andconcludes confidently that ldquoa new methodological situation inthe learning problem has developed where practical methodsare the result of a deep theoretical analysis of the statisticalbounds rather than the result of inventing new smart heuris-ticsrdquo Written still before the modern rehabilitation of neuralnetworks in the form of deep learning this now sounds overlyconfident The theme has indeed recently flared up again inthe community following Rahimi and Rechtrsquos speech at NIPS2017 in which they liken current machine learning research toalchemy

The predictable outrage notwithstanding Rahimi and Rechtnowhere commit to theory for the sake of theory they do nottake issue with the primacy of the pragmatic goal of findingalgorithms that work well in practice Their argument is thata theoretical understanding is important for this goal it is thestep back that facilitates so to speak a more efficient searchthrough the space of possible algorithms Continuous with thistheoretical work a further step back would be the foundationalwork that is the province of philosophers This concerns forinstance what it should mean when Rahimi and Recht speakabout ldquorigorous reliable verifiable knowledgerdquo and how thiscould be grounded in the relevant mathematics This also con-cerns the outer limits on learning algorithmsrsquo capabilities thesubject of my own work on universal prediction (University ofGroningen 2018)

The gap between theory and practice is particularly conspic-uous in what has been called a paradox of deep learning neural

97

networks perform astonishingly well in practice much betterthan they should in theory This is illustrated by Zhang et al(ldquoUnderstanding deep learning requires rethinking generaliza-tionrdquo ICLR 2017) with a simple experiment essentially theauthors took some standard architectures and data sets and thenchanged the data by reshuffling the labels in a random mannerThis guaranteed that there was no relation between instancesand labels yet the networks still managed to attain extremelylow training error entailing a wide divergence between train-ing error and generalization error (the latter must still be veryhigh the labeling is fully random so there is nothing to learn)But explanations based on statistical learning theory rely on aconnection between training error (or more generally the archi-tecture of the network) and generalization error wherefore thedemonstrated possibility of varying the latter while keeping theformer constant implies that these explanations simply do notapply Depending on onersquos perspective one could take this asanother illustration of the impotence of theory or of the direneed for a greater theoretical effort In either case the properway to understand deep learning is an important ongoing de-bate in the machine learning community that merits attentionfrom philosophers too

Tom SterkenburgMunich Centre for Mathematical Philosophy

Evidence-Based Medicine

Precision medicine is not new but it is currently a hot topic inEBM Often used interchangeably with personalised medicinethis new approach to healthcare aims to individualise treat-ments by recognizing the ways in which patients can differ per-son to person in the nature of disease and in patient responseto treatments It is well known that patients will respond dif-ferently to treatments that are effective for other people withthe same disease It is being increasingly recognised that thisis due to biological differences in the patients and the diseasesthemselves often at the genetic level Two very recent papersarticulate what precision medicine aims to do identify the ge-netic differences among patients with a certain disease and usenew technologies to enable us to make sense of the level ofcomplexity associated with using genetics to make predictions

One study classified subtypes of acute myeloid leukemia(AML) by their gene expression and DNA-methylation pat-terns AML is a single disease as it has a specific characteri-zation ldquoblocked myeloid lineage differentiation and accumu-lation of leukemic blast cellsrdquo But it is a ldquohighly heterogeneousdiseaserdquo as it can be caused by many different types of geneticalterations These different genetic alterations define separatesub-types of AML and patients with different sub-types willrequire different specific treatments This study identified pat-terns in the expression of genes in tumor cells from AML pa-tients which revealed gene regulatory networks that are spe-cific to the different sub-types of AML There were a numberof recurrent alterations important for lukemic growth identifiedthat present possibilities for targeted treatment Without thisknowledge patients may receive treatment that will not be ef-fective as it is not targeted against their specific AML subtype

The results of this study are claimed to be able to base tar-geted treatments for specific AML sub-types It is importantto note here that while precision medicine aims to take resultssuch as these and use them to target individual patients it is

actually sub-populations of patients with specific AML sub-types that can be targeted rather than individual patients per seDrugs will not be tailored to any one individual but novel drugscan be designed with specific sub-populations in mind and ex-tant drugs can be better targeted at specific sub-populationsThis is why precision medicine may be a better term than per-sonalised medicine as these techniques allow us to be moreprecise in targeting treatments to any one individual This mayseem like a small semantic difference but being clear about thistempers the ability to make grandiose claims about treating in-dividuals - it still might be the case that precision treatmentsare not effective for individuals due to chance or other factorsboth biological and environmental

Difficulties face the implementation of this research bothbroadly and specifically A difficulty facing precision medicineis practical in nature the level of complexity associated withusing results from disciplines such as genetics One way inwhich precision medicine hopes to overcome this challenge isthrough the use of technology in particular AI A practical dif-ficulty specific to oncology is that ldquotumors change over timeprogressing from benign to malignant becoming metastaticand developing resistance to certain therapiesrdquo Essentiallythe tumor cells evolve resulting in lsquointratumour heterogeneityrsquoThis means that due to tumour evolution using genetic mark-ers of cancer subtypes may quickly not be useful for targetingtherapies An enterprise hoping to provide solutions to bothdifficulties is the REVOLVER system that has recently pub-lished results on predicting changes in tumour evolution usingan AI system This system uses a machine-learning approachknown as transfer learning (TL)19 to predict the next steps inspecific tumor cell evolution Tumor evolution seems to be pre-dictable due to the prognostic value of histopathological stag-ing and molecular markers Problems facing using this data topredict tumor evolution are the complexity of patterns in theavailable data and stochastic forces in the mechanisms of tu-mor cell evolution The REVOLVER system can deal betterwith complexity and stochasticity than a human reasoner Thisenables more precise information for specific cancer patients asto what might happen during the course of their disease

I have written about AI in this column before and with thatcase (triaging in opthamology) the system was again used towork with large quantities of data to improve the efficiency ofhealth care delivery Precision medicine offers much in the wayof improving how we treat severe diseases but as the data setsget larger and more complex we will increasingly have to turnto AI It is promising that there is development in this fieldthat can meet this challenge Interestingly at present the datasets are not actually large enough for the AI systems to providethe high degree of precision we require But with precisionmedicine research considered a major growth area in medicinewe will likely not have to wait too long to overcome this prob-lem

DJ Auker-HowlettPhilosophy Kent

98

Events

December

KEiMS Conference on Knowledge Exchange in the Mathe-matical Sciences Aston University 3ndash4 DecemberCvC Causation vs Constitution Loosening the FrictionBergen 3ndash4 DecemberAxiMet Axiomatizing MetatheorymdashTruth Provability andBeyond Salzburg Austria 6ndash7 DecemberML4H Workshop on Machine Learning for Health MontrealCanada 8 DecemberDiI Workshop on Disagreement in Inquiry University of Tartu8 DecemberRLPO Reinforcement Learning Under Partial ObservabilityMontreal Canada 8 DecemberRPiB Real Patterns University of Barcelona 14 DecemberWrsquosBaD Whatrsquos so Bad About Dialetheism Kyoto Univer-sity Japan 15ndash17 December

January

PoMaL Graduate Conference on the Philosophy of Mathemat-ics and Logic University of Cambridge 19ndash20 JanuaryFiSS Foundations in Social SciencemdashMechanisms ActionsFunctions University of Duisburg-Essen Germany 24ndash25 Jan-uaryVoUiS Varieties of Understanding in Science Utrecht TheNetherlands 24ndash25 January

February

DMaMG Dark Matter and Modified Gravity ConferenceAachen Germany 6ndash8 February

Courses and Programmes

Courses

SSA Summer School on Argumentation Computational andLinguistic Perspectives on Argumentation Warsaw Poland 6ndash10 September

Programmes

APhil MAPhD in Analytic Philosophy University ofBarcelonaMaster Programme MA in Pure and Applied Logic Univer-sity of BarcelonaDoctoral Programme in Philosophy Language Mind andPractice Department of Philosophy University of ZurichSwitzerlandDoctoral Programme in Philosophy Department of Philoso-phy University of Milan ItalyLogiCS Joint doctoral program on Logical Methods in Com-puter Science TU Wien TU Graz and JKU Linz AustriaHPSM MA in the History and Philosophy of Science andMedicine Durham UniversityMaster Programme in Statistics University College DublinLoPhiSC Master in Logic Philosophy of Science and Epis-temology Pantheon-Sorbonne University (Paris 1) and Paris-Sorbonne University (Paris 4)Master Programme in Artificial Intelligence Radboud Uni-versity Nijmegen the NetherlandsMaster Programme Philosophy and Economics Institute ofPhilosophy University of BayreuthMA in Cognitive Science School of Politics InternationalStudies and Philosophy Queenrsquos University BelfastMA in Logic and the Philosophy ofMathematics Departmentof Philosophy University of BristolMA Programmes in Philosophy of Science University ofLeedsMA in Logic and Philosophy of Science Faculty of PhilosophyPhilosophy of Science and Study of Religion LMU MunichMA in Logic and Theory of Science Department of Logic ofthe Eotvos Lorand University Budapest HungaryMA in Metaphysics Language and Mind Department of Phi-losophy University of LiverpoolMA inMind Brain and Learning Westminster Institute of Ed-ucation Oxford Brookes UniversityMA in Philosophy by research Tilburg UniversityMA in Philosophy Science and Society TiLPS Tilburg Uni-versityMA in Philosophy of Biological and Cognitive Sciences De-partment of Philosophy University of BristolMA in Rhetoric School of Journalism Media and Communi-cation University of Central LancashireMA programmes in Philosophy of Language and Linguisticsand Philosophy of Mind and Psychology University of Birm-inghamMRes in Methods and Practices of Philosophical ResearchNorthern Institute of Philosophy University of AberdeenMSc in Applied Statistics Department of Economics Mathe-matics and Statistics Birkbeck University of LondonMSc in Applied Statistics and Datamining School of Mathe-matics and Statistics University of St Andrews

99

MSc in Artificial Intelligence Faculty of Engineering Uni-versity of LeedsMSc in Cognitiveamp Decision Sciences Psychology UniversityCollege LondonMSc in Cognitive Systems Language Learning and Reason-ing University of PotsdamMSc in Cognitive Science University of Osnabruck GermanyMSc in Cognitive PsychologyNeuropsychology School ofPsychology University of KentMSc in Logic Institute for Logic Language and ComputationUniversity of AmsterdamMSc in Mind Language amp Embodied Cognition School ofPhilosophy Psychology and Language Sciences University ofEdinburghMSc in Philosophy of Science Technology and Society Uni-versity of Twente The NetherlandsMRes in Cognitive Science and Humanities Language Com-munication and Organization Institute for Logic CognitionLanguage and Information University of the Basque Country(Donostia San Sebastian)OpenMind International School of Advanced Studies in Cog-nitive Sciences University of BucharestResearchMaster in Philosophy and Economics Erasmus Uni-versity Rotterdam The Netherlands

Jobs and Studentships

JobsAssistant Professor in Epistemology California State Uni-versity at Sacramento deadline until filledAssistant Professor in Logic amp Epistemology University ofNorth Carolina at Greensboro deadline until filledAssistant Professor in Philosophy of Medicine University ofNorth Carolina at Greensboro deadline until filledAssistant Professor in Philosophy of Science University ofFlorida deadline until filledLecturer in Actuarial Science University of California SantaBarbara deadline until filledProfessor in Philosophy of Science City University of NewYork deadline 3 DecemberResearch Associate in Bayesian modelling University ofSheffield deadline 12 DecemberResearch Fellow in Data Science University of Bristol dead-line 3 JanuaryLecturer in Statistical Science University College Londondeadline 9 February

100

  • Editorial
  • Features
  • News
  • Whats Hot in hellip
  • Events
  • Courses and Programmes
  • Jobs and Studentships

networks perform astonishingly well in practice much betterthan they should in theory This is illustrated by Zhang et al(ldquoUnderstanding deep learning requires rethinking generaliza-tionrdquo ICLR 2017) with a simple experiment essentially theauthors took some standard architectures and data sets and thenchanged the data by reshuffling the labels in a random mannerThis guaranteed that there was no relation between instancesand labels yet the networks still managed to attain extremelylow training error entailing a wide divergence between train-ing error and generalization error (the latter must still be veryhigh the labeling is fully random so there is nothing to learn)But explanations based on statistical learning theory rely on aconnection between training error (or more generally the archi-tecture of the network) and generalization error wherefore thedemonstrated possibility of varying the latter while keeping theformer constant implies that these explanations simply do notapply Depending on onersquos perspective one could take this asanother illustration of the impotence of theory or of the direneed for a greater theoretical effort In either case the properway to understand deep learning is an important ongoing de-bate in the machine learning community that merits attentionfrom philosophers too

Tom SterkenburgMunich Centre for Mathematical Philosophy

Evidence-Based Medicine

Precision medicine is not new but it is currently a hot topic inEBM Often used interchangeably with personalised medicinethis new approach to healthcare aims to individualise treat-ments by recognizing the ways in which patients can differ per-son to person in the nature of disease and in patient responseto treatments It is well known that patients will respond dif-ferently to treatments that are effective for other people withthe same disease It is being increasingly recognised that thisis due to biological differences in the patients and the diseasesthemselves often at the genetic level Two very recent papersarticulate what precision medicine aims to do identify the ge-netic differences among patients with a certain disease and usenew technologies to enable us to make sense of the level ofcomplexity associated with using genetics to make predictions

One study classified subtypes of acute myeloid leukemia(AML) by their gene expression and DNA-methylation pat-terns AML is a single disease as it has a specific characteri-zation ldquoblocked myeloid lineage differentiation and accumu-lation of leukemic blast cellsrdquo But it is a ldquohighly heterogeneousdiseaserdquo as it can be caused by many different types of geneticalterations These different genetic alterations define separatesub-types of AML and patients with different sub-types willrequire different specific treatments This study identified pat-terns in the expression of genes in tumor cells from AML pa-tients which revealed gene regulatory networks that are spe-cific to the different sub-types of AML There were a numberof recurrent alterations important for lukemic growth identifiedthat present possibilities for targeted treatment Without thisknowledge patients may receive treatment that will not be ef-fective as it is not targeted against their specific AML subtype

The results of this study are claimed to be able to base tar-geted treatments for specific AML sub-types It is importantto note here that while precision medicine aims to take resultssuch as these and use them to target individual patients it is

actually sub-populations of patients with specific AML sub-types that can be targeted rather than individual patients per seDrugs will not be tailored to any one individual but novel drugscan be designed with specific sub-populations in mind and ex-tant drugs can be better targeted at specific sub-populationsThis is why precision medicine may be a better term than per-sonalised medicine as these techniques allow us to be moreprecise in targeting treatments to any one individual This mayseem like a small semantic difference but being clear about thistempers the ability to make grandiose claims about treating in-dividuals - it still might be the case that precision treatmentsare not effective for individuals due to chance or other factorsboth biological and environmental

Difficulties face the implementation of this research bothbroadly and specifically A difficulty facing precision medicineis practical in nature the level of complexity associated withusing results from disciplines such as genetics One way inwhich precision medicine hopes to overcome this challenge isthrough the use of technology in particular AI A practical dif-ficulty specific to oncology is that ldquotumors change over timeprogressing from benign to malignant becoming metastaticand developing resistance to certain therapiesrdquo Essentiallythe tumor cells evolve resulting in lsquointratumour heterogeneityrsquoThis means that due to tumour evolution using genetic mark-ers of cancer subtypes may quickly not be useful for targetingtherapies An enterprise hoping to provide solutions to bothdifficulties is the REVOLVER system that has recently pub-lished results on predicting changes in tumour evolution usingan AI system This system uses a machine-learning approachknown as transfer learning (TL)19 to predict the next steps inspecific tumor cell evolution Tumor evolution seems to be pre-dictable due to the prognostic value of histopathological stag-ing and molecular markers Problems facing using this data topredict tumor evolution are the complexity of patterns in theavailable data and stochastic forces in the mechanisms of tu-mor cell evolution The REVOLVER system can deal betterwith complexity and stochasticity than a human reasoner Thisenables more precise information for specific cancer patients asto what might happen during the course of their disease

I have written about AI in this column before and with thatcase (triaging in opthamology) the system was again used towork with large quantities of data to improve the efficiency ofhealth care delivery Precision medicine offers much in the wayof improving how we treat severe diseases but as the data setsget larger and more complex we will increasingly have to turnto AI It is promising that there is development in this fieldthat can meet this challenge Interestingly at present the datasets are not actually large enough for the AI systems to providethe high degree of precision we require But with precisionmedicine research considered a major growth area in medicinewe will likely not have to wait too long to overcome this prob-lem

DJ Auker-HowlettPhilosophy Kent

98

Events

December

KEiMS Conference on Knowledge Exchange in the Mathe-matical Sciences Aston University 3ndash4 DecemberCvC Causation vs Constitution Loosening the FrictionBergen 3ndash4 DecemberAxiMet Axiomatizing MetatheorymdashTruth Provability andBeyond Salzburg Austria 6ndash7 DecemberML4H Workshop on Machine Learning for Health MontrealCanada 8 DecemberDiI Workshop on Disagreement in Inquiry University of Tartu8 DecemberRLPO Reinforcement Learning Under Partial ObservabilityMontreal Canada 8 DecemberRPiB Real Patterns University of Barcelona 14 DecemberWrsquosBaD Whatrsquos so Bad About Dialetheism Kyoto Univer-sity Japan 15ndash17 December

January

PoMaL Graduate Conference on the Philosophy of Mathemat-ics and Logic University of Cambridge 19ndash20 JanuaryFiSS Foundations in Social SciencemdashMechanisms ActionsFunctions University of Duisburg-Essen Germany 24ndash25 Jan-uaryVoUiS Varieties of Understanding in Science Utrecht TheNetherlands 24ndash25 January

February

DMaMG Dark Matter and Modified Gravity ConferenceAachen Germany 6ndash8 February

Courses and Programmes

Courses

SSA Summer School on Argumentation Computational andLinguistic Perspectives on Argumentation Warsaw Poland 6ndash10 September

Programmes

APhil MAPhD in Analytic Philosophy University ofBarcelonaMaster Programme MA in Pure and Applied Logic Univer-sity of BarcelonaDoctoral Programme in Philosophy Language Mind andPractice Department of Philosophy University of ZurichSwitzerlandDoctoral Programme in Philosophy Department of Philoso-phy University of Milan ItalyLogiCS Joint doctoral program on Logical Methods in Com-puter Science TU Wien TU Graz and JKU Linz AustriaHPSM MA in the History and Philosophy of Science andMedicine Durham UniversityMaster Programme in Statistics University College DublinLoPhiSC Master in Logic Philosophy of Science and Epis-temology Pantheon-Sorbonne University (Paris 1) and Paris-Sorbonne University (Paris 4)Master Programme in Artificial Intelligence Radboud Uni-versity Nijmegen the NetherlandsMaster Programme Philosophy and Economics Institute ofPhilosophy University of BayreuthMA in Cognitive Science School of Politics InternationalStudies and Philosophy Queenrsquos University BelfastMA in Logic and the Philosophy ofMathematics Departmentof Philosophy University of BristolMA Programmes in Philosophy of Science University ofLeedsMA in Logic and Philosophy of Science Faculty of PhilosophyPhilosophy of Science and Study of Religion LMU MunichMA in Logic and Theory of Science Department of Logic ofthe Eotvos Lorand University Budapest HungaryMA in Metaphysics Language and Mind Department of Phi-losophy University of LiverpoolMA inMind Brain and Learning Westminster Institute of Ed-ucation Oxford Brookes UniversityMA in Philosophy by research Tilburg UniversityMA in Philosophy Science and Society TiLPS Tilburg Uni-versityMA in Philosophy of Biological and Cognitive Sciences De-partment of Philosophy University of BristolMA in Rhetoric School of Journalism Media and Communi-cation University of Central LancashireMA programmes in Philosophy of Language and Linguisticsand Philosophy of Mind and Psychology University of Birm-inghamMRes in Methods and Practices of Philosophical ResearchNorthern Institute of Philosophy University of AberdeenMSc in Applied Statistics Department of Economics Mathe-matics and Statistics Birkbeck University of LondonMSc in Applied Statistics and Datamining School of Mathe-matics and Statistics University of St Andrews

99

MSc in Artificial Intelligence Faculty of Engineering Uni-versity of LeedsMSc in Cognitiveamp Decision Sciences Psychology UniversityCollege LondonMSc in Cognitive Systems Language Learning and Reason-ing University of PotsdamMSc in Cognitive Science University of Osnabruck GermanyMSc in Cognitive PsychologyNeuropsychology School ofPsychology University of KentMSc in Logic Institute for Logic Language and ComputationUniversity of AmsterdamMSc in Mind Language amp Embodied Cognition School ofPhilosophy Psychology and Language Sciences University ofEdinburghMSc in Philosophy of Science Technology and Society Uni-versity of Twente The NetherlandsMRes in Cognitive Science and Humanities Language Com-munication and Organization Institute for Logic CognitionLanguage and Information University of the Basque Country(Donostia San Sebastian)OpenMind International School of Advanced Studies in Cog-nitive Sciences University of BucharestResearchMaster in Philosophy and Economics Erasmus Uni-versity Rotterdam The Netherlands

Jobs and Studentships

JobsAssistant Professor in Epistemology California State Uni-versity at Sacramento deadline until filledAssistant Professor in Logic amp Epistemology University ofNorth Carolina at Greensboro deadline until filledAssistant Professor in Philosophy of Medicine University ofNorth Carolina at Greensboro deadline until filledAssistant Professor in Philosophy of Science University ofFlorida deadline until filledLecturer in Actuarial Science University of California SantaBarbara deadline until filledProfessor in Philosophy of Science City University of NewYork deadline 3 DecemberResearch Associate in Bayesian modelling University ofSheffield deadline 12 DecemberResearch Fellow in Data Science University of Bristol dead-line 3 JanuaryLecturer in Statistical Science University College Londondeadline 9 February

100

  • Editorial
  • Features
  • News
  • Whats Hot in hellip
  • Events
  • Courses and Programmes
  • Jobs and Studentships

Events

December

KEiMS Conference on Knowledge Exchange in the Mathe-matical Sciences Aston University 3ndash4 DecemberCvC Causation vs Constitution Loosening the FrictionBergen 3ndash4 DecemberAxiMet Axiomatizing MetatheorymdashTruth Provability andBeyond Salzburg Austria 6ndash7 DecemberML4H Workshop on Machine Learning for Health MontrealCanada 8 DecemberDiI Workshop on Disagreement in Inquiry University of Tartu8 DecemberRLPO Reinforcement Learning Under Partial ObservabilityMontreal Canada 8 DecemberRPiB Real Patterns University of Barcelona 14 DecemberWrsquosBaD Whatrsquos so Bad About Dialetheism Kyoto Univer-sity Japan 15ndash17 December

January

PoMaL Graduate Conference on the Philosophy of Mathemat-ics and Logic University of Cambridge 19ndash20 JanuaryFiSS Foundations in Social SciencemdashMechanisms ActionsFunctions University of Duisburg-Essen Germany 24ndash25 Jan-uaryVoUiS Varieties of Understanding in Science Utrecht TheNetherlands 24ndash25 January

February

DMaMG Dark Matter and Modified Gravity ConferenceAachen Germany 6ndash8 February

Courses and Programmes

Courses

SSA Summer School on Argumentation Computational andLinguistic Perspectives on Argumentation Warsaw Poland 6ndash10 September

Programmes

APhil MAPhD in Analytic Philosophy University ofBarcelonaMaster Programme MA in Pure and Applied Logic Univer-sity of BarcelonaDoctoral Programme in Philosophy Language Mind andPractice Department of Philosophy University of ZurichSwitzerlandDoctoral Programme in Philosophy Department of Philoso-phy University of Milan ItalyLogiCS Joint doctoral program on Logical Methods in Com-puter Science TU Wien TU Graz and JKU Linz AustriaHPSM MA in the History and Philosophy of Science andMedicine Durham UniversityMaster Programme in Statistics University College DublinLoPhiSC Master in Logic Philosophy of Science and Epis-temology Pantheon-Sorbonne University (Paris 1) and Paris-Sorbonne University (Paris 4)Master Programme in Artificial Intelligence Radboud Uni-versity Nijmegen the NetherlandsMaster Programme Philosophy and Economics Institute ofPhilosophy University of BayreuthMA in Cognitive Science School of Politics InternationalStudies and Philosophy Queenrsquos University BelfastMA in Logic and the Philosophy ofMathematics Departmentof Philosophy University of BristolMA Programmes in Philosophy of Science University ofLeedsMA in Logic and Philosophy of Science Faculty of PhilosophyPhilosophy of Science and Study of Religion LMU MunichMA in Logic and Theory of Science Department of Logic ofthe Eotvos Lorand University Budapest HungaryMA in Metaphysics Language and Mind Department of Phi-losophy University of LiverpoolMA inMind Brain and Learning Westminster Institute of Ed-ucation Oxford Brookes UniversityMA in Philosophy by research Tilburg UniversityMA in Philosophy Science and Society TiLPS Tilburg Uni-versityMA in Philosophy of Biological and Cognitive Sciences De-partment of Philosophy University of BristolMA in Rhetoric School of Journalism Media and Communi-cation University of Central LancashireMA programmes in Philosophy of Language and Linguisticsand Philosophy of Mind and Psychology University of Birm-inghamMRes in Methods and Practices of Philosophical ResearchNorthern Institute of Philosophy University of AberdeenMSc in Applied Statistics Department of Economics Mathe-matics and Statistics Birkbeck University of LondonMSc in Applied Statistics and Datamining School of Mathe-matics and Statistics University of St Andrews

99

MSc in Artificial Intelligence Faculty of Engineering Uni-versity of LeedsMSc in Cognitiveamp Decision Sciences Psychology UniversityCollege LondonMSc in Cognitive Systems Language Learning and Reason-ing University of PotsdamMSc in Cognitive Science University of Osnabruck GermanyMSc in Cognitive PsychologyNeuropsychology School ofPsychology University of KentMSc in Logic Institute for Logic Language and ComputationUniversity of AmsterdamMSc in Mind Language amp Embodied Cognition School ofPhilosophy Psychology and Language Sciences University ofEdinburghMSc in Philosophy of Science Technology and Society Uni-versity of Twente The NetherlandsMRes in Cognitive Science and Humanities Language Com-munication and Organization Institute for Logic CognitionLanguage and Information University of the Basque Country(Donostia San Sebastian)OpenMind International School of Advanced Studies in Cog-nitive Sciences University of BucharestResearchMaster in Philosophy and Economics Erasmus Uni-versity Rotterdam The Netherlands

Jobs and Studentships

JobsAssistant Professor in Epistemology California State Uni-versity at Sacramento deadline until filledAssistant Professor in Logic amp Epistemology University ofNorth Carolina at Greensboro deadline until filledAssistant Professor in Philosophy of Medicine University ofNorth Carolina at Greensboro deadline until filledAssistant Professor in Philosophy of Science University ofFlorida deadline until filledLecturer in Actuarial Science University of California SantaBarbara deadline until filledProfessor in Philosophy of Science City University of NewYork deadline 3 DecemberResearch Associate in Bayesian modelling University ofSheffield deadline 12 DecemberResearch Fellow in Data Science University of Bristol dead-line 3 JanuaryLecturer in Statistical Science University College Londondeadline 9 February

100

  • Editorial
  • Features
  • News
  • Whats Hot in hellip
  • Events
  • Courses and Programmes
  • Jobs and Studentships

MSc in Artificial Intelligence Faculty of Engineering Uni-versity of LeedsMSc in Cognitiveamp Decision Sciences Psychology UniversityCollege LondonMSc in Cognitive Systems Language Learning and Reason-ing University of PotsdamMSc in Cognitive Science University of Osnabruck GermanyMSc in Cognitive PsychologyNeuropsychology School ofPsychology University of KentMSc in Logic Institute for Logic Language and ComputationUniversity of AmsterdamMSc in Mind Language amp Embodied Cognition School ofPhilosophy Psychology and Language Sciences University ofEdinburghMSc in Philosophy of Science Technology and Society Uni-versity of Twente The NetherlandsMRes in Cognitive Science and Humanities Language Com-munication and Organization Institute for Logic CognitionLanguage and Information University of the Basque Country(Donostia San Sebastian)OpenMind International School of Advanced Studies in Cog-nitive Sciences University of BucharestResearchMaster in Philosophy and Economics Erasmus Uni-versity Rotterdam The Netherlands

Jobs and Studentships

JobsAssistant Professor in Epistemology California State Uni-versity at Sacramento deadline until filledAssistant Professor in Logic amp Epistemology University ofNorth Carolina at Greensboro deadline until filledAssistant Professor in Philosophy of Medicine University ofNorth Carolina at Greensboro deadline until filledAssistant Professor in Philosophy of Science University ofFlorida deadline until filledLecturer in Actuarial Science University of California SantaBarbara deadline until filledProfessor in Philosophy of Science City University of NewYork deadline 3 DecemberResearch Associate in Bayesian modelling University ofSheffield deadline 12 DecemberResearch Fellow in Data Science University of Bristol dead-line 3 JanuaryLecturer in Statistical Science University College Londondeadline 9 February

100

  • Editorial
  • Features
  • News
  • Whats Hot in hellip
  • Events
  • Courses and Programmes
  • Jobs and Studentships