Seminar by Prof Bruce Bassett at IAP, Paris, October 2013
-
Upload
cosmoaims-bassett -
Category
Education
-
view
877 -
download
3
description
Transcript of Seminar by Prof Bruce Bassett at IAP, Paris, October 2013
![Page 1: Seminar by Prof Bruce Bassett at IAP, Paris, October 2013](https://reader033.fdocuments.in/reader033/viewer/2022061219/54ba27304a79591f4f8b4581/html5/thumbnails/1.jpg)
Rise of the Machines
Bruce Bassett AIMS – SAAO – UCT
Meudon
IAP – 4 October 2013
![Page 2: Seminar by Prof Bruce Bassett at IAP, Paris, October 2013](https://reader033.fdocuments.in/reader033/viewer/2022061219/54ba27304a79591f4f8b4581/html5/thumbnails/2.jpg)
www.aims.ac.za
![Page 3: Seminar by Prof Bruce Bassett at IAP, Paris, October 2013](https://reader033.fdocuments.in/reader033/viewer/2022061219/54ba27304a79591f4f8b4581/html5/thumbnails/3.jpg)
Why am I interested in this topic?
![Page 4: Seminar by Prof Bruce Bassett at IAP, Paris, October 2013](https://reader033.fdocuments.in/reader033/viewer/2022061219/54ba27304a79591f4f8b4581/html5/thumbnails/4.jpg)
1Drinking from the Fire Hydrant
• SKA, LSST, CTA, EUCLID, ALMA, LOFAR, BigBOSS, DESI, etc…
• Will produce around an Exabyte a day. Humanity produced about 10 exabytes prior to 2000
• This is not more of the same. No one can really prepare our students for this amount of data, because no one (especially in astronomy) has ever seen it before.
![Page 5: Seminar by Prof Bruce Bassett at IAP, Paris, October 2013](https://reader033.fdocuments.in/reader033/viewer/2022061219/54ba27304a79591f4f8b4581/html5/thumbnails/5.jpg)
2
![Page 6: Seminar by Prof Bruce Bassett at IAP, Paris, October 2013](https://reader033.fdocuments.in/reader033/viewer/2022061219/54ba27304a79591f4f8b4581/html5/thumbnails/6.jpg)
It raises fascinating fundamental issues in articifical intelligence, human cognition, the nature of science etc…
3
![Page 7: Seminar by Prof Bruce Bassett at IAP, Paris, October 2013](https://reader033.fdocuments.in/reader033/viewer/2022061219/54ba27304a79591f4f8b4581/html5/thumbnails/7.jpg)
A typical Astronomy example…
Given repeat imaging of the sky:
(a) identify plausible supernova candidates and (b) distinguish Type Ia’s from non-Ia’s using
multi-band light-curves…
![Page 8: Seminar by Prof Bruce Bassett at IAP, Paris, October 2013](https://reader033.fdocuments.in/reader033/viewer/2022061219/54ba27304a79591f4f8b4581/html5/thumbnails/8.jpg)
![Page 9: Seminar by Prof Bruce Bassett at IAP, Paris, October 2013](https://reader033.fdocuments.in/reader033/viewer/2022061219/54ba27304a79591f4f8b4581/html5/thumbnails/9.jpg)
![Page 10: Seminar by Prof Bruce Bassett at IAP, Paris, October 2013](https://reader033.fdocuments.in/reader033/viewer/2022061219/54ba27304a79591f4f8b4581/html5/thumbnails/10.jpg)
![Page 11: Seminar by Prof Bruce Bassett at IAP, Paris, October 2013](https://reader033.fdocuments.in/reader033/viewer/2022061219/54ba27304a79591f4f8b4581/html5/thumbnails/11.jpg)
![Page 12: Seminar by Prof Bruce Bassett at IAP, Paris, October 2013](https://reader033.fdocuments.in/reader033/viewer/2022061219/54ba27304a79591f4f8b4581/html5/thumbnails/12.jpg)
![Page 13: Seminar by Prof Bruce Bassett at IAP, Paris, October 2013](https://reader033.fdocuments.in/reader033/viewer/2022061219/54ba27304a79591f4f8b4581/html5/thumbnails/13.jpg)
Feature extraction using Eigenimages
du Buisson, BB et al 2013
![Page 14: Seminar by Prof Bruce Bassett at IAP, Paris, October 2013](https://reader033.fdocuments.in/reader033/viewer/2022061219/54ba27304a79591f4f8b4581/html5/thumbnails/14.jpg)
Classification…
![Page 15: Seminar by Prof Bruce Bassett at IAP, Paris, October 2013](https://reader033.fdocuments.in/reader033/viewer/2022061219/54ba27304a79591f4f8b4581/html5/thumbnails/15.jpg)
Example: Boosting
An ensemble classifier based on trees…
Combining many quasi-independent classifiers gives a better classifier
(democracy should allow better decision making)
![Page 16: Seminar by Prof Bruce Bassett at IAP, Paris, October 2013](https://reader033.fdocuments.in/reader033/viewer/2022061219/54ba27304a79591f4f8b4581/html5/thumbnails/16.jpg)
Newling
![Page 17: Seminar by Prof Bruce Bassett at IAP, Paris, October 2013](https://reader033.fdocuments.in/reader033/viewer/2022061219/54ba27304a79591f4f8b4581/html5/thumbnails/17.jpg)
Newling
![Page 18: Seminar by Prof Bruce Bassett at IAP, Paris, October 2013](https://reader033.fdocuments.in/reader033/viewer/2022061219/54ba27304a79591f4f8b4581/html5/thumbnails/18.jpg)
Newling
![Page 19: Seminar by Prof Bruce Bassett at IAP, Paris, October 2013](https://reader033.fdocuments.in/reader033/viewer/2022061219/54ba27304a79591f4f8b4581/html5/thumbnails/19.jpg)
![Page 20: Seminar by Prof Bruce Bassett at IAP, Paris, October 2013](https://reader033.fdocuments.in/reader033/viewer/2022061219/54ba27304a79591f4f8b4581/html5/thumbnails/20.jpg)
Representative?
Newling et al
How do we trust machine learning results?
![Page 21: Seminar by Prof Bruce Bassett at IAP, Paris, October 2013](https://reader033.fdocuments.in/reader033/viewer/2022061219/54ba27304a79591f4f8b4581/html5/thumbnails/21.jpg)
![Page 22: Seminar by Prof Bruce Bassett at IAP, Paris, October 2013](https://reader033.fdocuments.in/reader033/viewer/2022061219/54ba27304a79591f4f8b4581/html5/thumbnails/22.jpg)
There is now a better way. Petabytes allow us to say: "Correlation is enough." We can stop looking for models. We can analyze the data without hypotheses about what it might show. We can throw the numbers into the biggest computing clusters the world has ever seen and let statistical algorithms find patterns where science cannot.
“
”
![Page 23: Seminar by Prof Bruce Bassett at IAP, Paris, October 2013](https://reader033.fdocuments.in/reader033/viewer/2022061219/54ba27304a79591f4f8b4581/html5/thumbnails/23.jpg)
Some examples…
Discovering Newton’s Laws with symbolic regression…
"Distilling Free-Form Natural Laws from Experimental Data.” Schmidt and Lipson. Science, 2009
![Page 24: Seminar by Prof Bruce Bassett at IAP, Paris, October 2013](https://reader033.fdocuments.in/reader033/viewer/2022061219/54ba27304a79591f4f8b4581/html5/thumbnails/24.jpg)
Bruce Bassett
ADAM
“The first non-human contribution to human Knowledge”
![Page 25: Seminar by Prof Bruce Bassett at IAP, Paris, October 2013](https://reader033.fdocuments.in/reader033/viewer/2022061219/54ba27304a79591f4f8b4581/html5/thumbnails/25.jpg)
But is this really what we mean by (great) science?
![Page 26: Seminar by Prof Bruce Bassett at IAP, Paris, October 2013](https://reader033.fdocuments.in/reader033/viewer/2022061219/54ba27304a79591f4f8b4581/html5/thumbnails/26.jpg)
Not really
• Because great science involves two things:
– Finding the right representation/parameterisation (the “genius” part)
– Finding the right parameters within the representation (the “algorithmic” part)
Example: Discovering GR from the perehelion shift of mercury…already known to be a problem by 1859.
![Page 27: Seminar by Prof Bruce Bassett at IAP, Paris, October 2013](https://reader033.fdocuments.in/reader033/viewer/2022061219/54ba27304a79591f4f8b4581/html5/thumbnails/27.jpg)
Non-representative Training Set!
![Page 28: Seminar by Prof Bruce Bassett at IAP, Paris, October 2013](https://reader033.fdocuments.in/reader033/viewer/2022061219/54ba27304a79591f4f8b4581/html5/thumbnails/28.jpg)
The first claim of a sudden transition in w(z) - 2002
BB, Kunz, Silk, Ungarelli 2002
![Page 29: Seminar by Prof Bruce Bassett at IAP, Paris, October 2013](https://reader033.fdocuments.in/reader033/viewer/2022061219/54ba27304a79591f4f8b4581/html5/thumbnails/29.jpg)
The first claim of a sudden transition in w(z) - 2002
We assumed ns = 1 and t = 0
![Page 30: Seminar by Prof Bruce Bassett at IAP, Paris, October 2013](https://reader033.fdocuments.in/reader033/viewer/2022061219/54ba27304a79591f4f8b4581/html5/thumbnails/30.jpg)
Why did we do it?
• Because everyone did it at the time (bad training set!)
• Because we did grid computations, even though MCMC had been published a couple of years before.
• We made the mistake of not looking at what might cause the same effect…
• A machine learning algorithm might not have made the same mistake…
![Page 31: Seminar by Prof Bruce Bassett at IAP, Paris, October 2013](https://reader033.fdocuments.in/reader033/viewer/2022061219/54ba27304a79591f4f8b4581/html5/thumbnails/31.jpg)
To do “real” science with machine learning one would need a universal space for describing general theories that one could then search through algorithmically, removing the need for “genius”...
![Page 32: Seminar by Prof Bruce Bassett at IAP, Paris, October 2013](https://reader033.fdocuments.in/reader033/viewer/2022061219/54ba27304a79591f4f8b4581/html5/thumbnails/32.jpg)
Historical ideas about the perfect language
• 1200’s – Raymond Lull
• 1600’s Leibniz and the Characterisitca Universalis which lead to calculus. Unpopular at the time
because it seemed to remove the need for creativity…
![Page 33: Seminar by Prof Bruce Bassett at IAP, Paris, October 2013](https://reader033.fdocuments.in/reader033/viewer/2022061219/54ba27304a79591f4f8b4581/html5/thumbnails/33.jpg)
We have spoken of the art of complication of the sciences, i.e., of inventive logic... But when the tables of categories of our art of complication have been formed, something greater will emerge. For let the first terms, of the combination of which all others consist, be designated by signs; these signs will be a kind of alphabet.
- Leibniz 1666 “The art of combination.”
“
”
![Page 34: Seminar by Prof Bruce Bassett at IAP, Paris, October 2013](https://reader033.fdocuments.in/reader033/viewer/2022061219/54ba27304a79591f4f8b4581/html5/thumbnails/34.jpg)
Hilbert c.1900: formal axiomatic theory
Goedel/Turing: Not possible – incompleteness, uncomputability.
“This statement is unprovable”
The Halting Problem
![Page 35: Seminar by Prof Bruce Bassett at IAP, Paris, October 2013](https://reader033.fdocuments.in/reader033/viewer/2022061219/54ba27304a79591f4f8b4581/html5/thumbnails/35.jpg)
Can physics be fundamentally described only in terms of computable objects?
Enumeration of 4d topologies…
![Page 36: Seminar by Prof Bruce Bassett at IAP, Paris, October 2013](https://reader033.fdocuments.in/reader033/viewer/2022061219/54ba27304a79591f4f8b4581/html5/thumbnails/36.jpg)
Despite this, today there is very active work on automated theorem generation and automated reasoning...
Combine this with numerical methods and machine learning and it might soon be possible to have fully automated workflows.
We joke about computers writing scientific papers for us, but perhaps it is us who need to learn to write papers “properly”?
![Page 37: Seminar by Prof Bruce Bassett at IAP, Paris, October 2013](https://reader033.fdocuments.in/reader033/viewer/2022061219/54ba27304a79591f4f8b4581/html5/thumbnails/37.jpg)
• Arguably the best-written papers are the ones that most closely follow the ideal of
{axioms, assumptions} + {data} + {derived propositions} results
• We try to emulate a digital/algorithmic ideal in our papers. i.e. we try to emulate computers (modulo the explanations and analogies used to explain our thinking)
![Page 38: Seminar by Prof Bruce Bassett at IAP, Paris, October 2013](https://reader033.fdocuments.in/reader033/viewer/2022061219/54ba27304a79591f4f8b4581/html5/thumbnails/38.jpg)
In 1900 this is what it meant to be a “computer” in astronomy:
Perhaps by 2030 the meaning of “computer” will have changed dramatically again…
http://blogs.smithsonianmag.com/