d-VMP: Distributed Variational Message Passingalessandro/pgm/Martinez.pdf · d-VMP: Distributed...
Transcript of d-VMP: Distributed Variational Message Passingalessandro/pgm/Martinez.pdf · d-VMP: Distributed...
d-VMP:
Distributed Variational Message Passing
Andrés R. Masegosa1, Ana M. Martínez2,Helge Langseth1, Thomas D. Nielsen2, Antonio Salmerón3,
Darío Ramos-López3, Anders L. Madsen2,4
1Department of Computer Science, Aalborg University, Denmark2 Department of Computer and Information Science,
The Norwegian University of Science and Technology, Norway3Department of Mathematics, University of Almería, Spain
4 Hugin Expert A/S, Aalborg, Denmark
Int. Conf. on Probabilistic Graphical Models, Lugano, Sept. 6–9, 2016 1
Outline
1 Motivation
2 Variational Message Passing
3 d-VMP
4 Experimental results
5 Conclusions
Int. Conf. on Probabilistic Graphical Models, Lugano, Sept. 6–9, 2016 2
Outline
1 Motivation
2 Variational Message Passing
3 d-VMP
4 Experimental results
5 Conclusions
Int. Conf. on Probabilistic Graphical Models, Lugano, Sept. 6–9, 2016 3
Motivation
I LargeI ImbalanceI ? valuesI Complex distributions
Int. Conf. on Probabilistic Graphical Models, Lugano, Sept. 6–9, 2016 4
Motivation
I LargeI ImbalanceI ? valuesI Complex distributions
Attribute range
Den
sity
0 1 2 3 4 5 6
01
23
45
6
●
●
●
●
SVI 1%SVI 5%SVI 10%VMP 1%
Int. Conf. on Probabilistic Graphical Models, Lugano, Sept. 6–9, 2016 5
Motivation
I Goal: learn a generative model for a finantial dataset to monitor thecustomers and make predictions for a single customer.
X i H i
✓↵
i = 1, . . . ,N
Int. Conf. on Probabilistic Graphical Models, Lugano, Sept. 6–9, 2016 6
Popular existing approach: SVI
IStochastic Variational Inference: iteratively updates the modelparameters based on subsampled data batches.
INo estimation of all local hidden variables of the model.
INo generation of lower bound.
IPoor fit if batch of data is not representative from all data.
Int. Conf. on Probabilistic Graphical Models, Lugano, Sept. 6–9, 2016 7
Our contribution:
Id-VMP: a distributed message passing scheme.
IDefined for a broader class of models (than SVI).
IBetter and faster convergence results compared to SVI.
IPosterior over all local latent variables and lower bound.
Int. Conf. on Probabilistic Graphical Models, Lugano, Sept. 6–9, 2016 8
Outline
1 Motivation
2 Variational Message Passing
3 d-VMP
4 Experimental results
5 Conclusions
Int. Conf. on Probabilistic Graphical Models, Lugano, Sept. 6–9, 2016 9
Models:
I Bayesian learning on iid. data using conjugate exponential BN models:
ln p(X ) = ln hX + sX · ⌘ � AX (⌘)
X i H i
✓↵
i = 1, . . . ,N
I We want to calculate p(✓,H |D).
Int. Conf. on Probabilistic Graphical Models, Lugano, Sept. 6–9, 2016 10
Variational Inference:
I Approximate p(✓,H |D) (often intractable) by finding tractableposterior distributions q 2 Q by minimizing:
minq(✓,H)2Q
KL(q(✓,H)|p(✓,H |D)),
I In the mean field variational approach, Q is assumed to fullyfactorize:
q(✓,H) =MY
k=1
q(✓k)NY
i=1
JY
j=1
q(Hi ,j),
Int. Conf. on Probabilistic Graphical Models, Lugano, Sept. 6–9, 2016 11
Variational Inference:
I Approximate p(✓,H |D) (often intractable) by finding tractableposterior distributions q 2 Q by minimizing:
minq(✓,H)2Q
KL(q(✓,H)|p(✓,H |D)),
I In the mean field variational approach, Q is assumed to fullyfactorize:
q(✓,H) =MY
k=1
q(✓k)NY
i=1
JY
j=1
q(Hi ,j),
Int. Conf. on Probabilistic Graphical Models, Lugano, Sept. 6–9, 2016 12
Variational Inference:
I Variational Inference exploits:
lnP(D)
constant
= L(q(✓,H))
Maximize
+ KL(q(✓,H)|p(✓,H |D))
Minimize
,
I Iterative coordinate ascent of the variational distributions.I Updates in the variational distribution of a variable only
involves variables in its Markov blanket.I Coordinate ascent algorithm formulated as a message passing
scheme.
Int. Conf. on Probabilistic Graphical Models, Lugano, Sept. 6–9, 2016 13
Variational Message Passing, VMP:
IMessage from parent to child: moment parameters(expectation of the sufficient statistics).
IMessage from child to parent: natural parameters(based on the messages received from the co-parents).
Int. Conf. on Probabilistic Graphical Models, Lugano, Sept. 6–9, 2016 14
Outline
1 Motivation
2 Variational Message Passing
3 d-VMP
4 Experimental results
5 Conclusions
Int. Conf. on Probabilistic Graphical Models, Lugano, Sept. 6–9, 2016 15
Distributed optimization of the lower bound:
Master
✓↵
q(t)(✓)
X
H
✓1
i = 1, . . . ,N
Slave 1
q(t)(H1)
X
H
✓2
i = 1, . . . ,N
Slave 2
q(t)(H2)
X
H
✓3
i = 1, . . . ,N
Slave 3
q(t)(H3)
q(t)(✓) is broadcasted to all the slave nodes.
maxq2Q L(q(✓,H))
Int. Conf. on Probabilistic Graphical Models, Lugano, Sept. 6–9, 2016 16
Distributed optimization of the lower bound:
Master
✓↵
q(t)(✓)
X
H
✓1
i = 1, . . . ,N
Slave 1
q(t)(H1)
X
H
✓2
i = 1, . . . ,N
Slave 2
q(t)(H2)
X
H
✓3
i = 1, . . . ,N
Slave 3
q(t)(H3)
q(t+1)(H) = arg maxq(H) L(q(H), q(t)(✓))
maxq2Q L(q(✓,H))
Int. Conf. on Probabilistic Graphical Models, Lugano, Sept. 6–9, 2016 17
Distributed optimization of the lower bound:
Master
✓↵
q(t)(✓)
X
H
✓1
i = 1, . . . ,N
Slave 1
q(t)(H1)
X
H
✓2
i = 1, . . . ,N
Slave 2
q(t)(H2)
X
H
✓3
i = 1, . . . ,N
Slave 3
q(t)(H3)
q(t+1)(Hn) = arg maxq(Hn) Ln(q(Hn), q(t)(✓))
maxq2Q L(q(✓,H))
Int. Conf. on Probabilistic Graphical Models, Lugano, Sept. 6–9, 2016 18
Distributed optimization of the lower bound:
Master
✓↵
q(t)(✓)
X
H
✓1
i = 1, . . . ,N
Slave 1
q(t)(H1)
X
H
✓2
i = 1, . . . ,N
Slave 2
q(t)(H2)
X
H
✓3
i = 1, . . . ,N
Slave 3
q(t)(H3)
q(t+1)(✓) = arg maxq(✓)
L(q(t)(H), q(✓))
May be coupled
maxq2Q L(q(✓,H))
Int. Conf. on Probabilistic Graphical Models, Lugano, Sept. 6–9, 2016 19
Candidate solutions:
I Resort to a generalized mean-field approximation as SVI: doesnot factorize over the global parameters.
IProhibitive for models with a large number of global (coupled)
parameters, e.g. linear regression.
I Our proposal: VMP as a distributed projected natural
gradient ascent algorithm (PGNA).
Int. Conf. on Probabilistic Graphical Models, Lugano, Sept. 6–9, 2016 20
d-VMP as a projected natural gradient ascent
IInsight 1: VMP can be expressed as a projected naturalgradient ascent algorithm.
⌘(t+1)X = ⌘(t)
X + ⇢X ,t [r⌘L(⌘(t))]+X (1)
I [·] is the projection operator.
Int. Conf. on Probabilistic Graphical Models, Lugano, Sept. 6–9, 2016 21
d-VMP as a projected natural gradient ascent
IInsight 2: The natural gradient of the lower bound can beexpressed as follows:
r⌘✓L = mPa(✓)!✓ +
XmHi!✓
IThe gradient can be computed in parallel.
Int. Conf. on Probabilistic Graphical Models, Lugano, Sept. 6–9, 2016 22
d-VMP as a projected natural gradient ascent
IInsight 3: Global parameters are “coupled” only if they belongto each other’s Markov blanket.
IDefine a disjoint partition of the global parameters:
R = {J1, . . . ,JS}
Int. Conf. on Probabilistic Graphical Models, Lugano, Sept. 6–9, 2016 23
d-VMP as a projected natural gradient ascent
I d-VMP is based on performing independent global updatesover the global parameters of each partition:
⌘(t+1)Jr
= ⌘(t)Jr
+ ⇢r ,t [r⌘L(⌘(t))]+Jr
I ⇢r ,t is the learning rate. If |Jr | = 1 then ⇢r ,t = 1.
Int. Conf. on Probabilistic Graphical Models, Lugano, Sept. 6–9, 2016 24
dVMP as a distributed PNGA algorithm:
Master
✓↵
⌘(t+1)Jr
X
H
✓1
i = 1, . . . ,N
Slave 1
⌘(t+1)1
X
H
✓2
i = 1, . . . ,N
Slave 2
⌘(t+1)2
X
H
✓3
i = 1, . . . ,N
Slave 3
⌘(t+1)3
⌘(t)Jr+ ⇢r ,t [O⌘L(⌘(t))]+Jr
For all Jr 2 R
Int. Conf. on Probabilistic Graphical Models, Lugano, Sept. 6–9, 2016 25
Outline
1 Motivation
2 Variational Message Passing
3 d-VMP
4 Experimental results
5 Conclusions
Int. Conf. on Probabilistic Graphical Models, Lugano, Sept. 6–9, 2016 26
Model fit to the data
X ij H ijYi
Hi ✓
↵
j = 1, . . . , J
i = 1, . . . ,N
Attribute rangeD
ensi
ty
0 1 2 3 4 5 6
01
23
45
6
●
●
●
●
SVI 1%SVI 5%SVI 10%VMP 1%
I Representative sample of 55K clients (N) and 33 attributes (J).I “Unrolled” model of more than 3.5M nodes (75% latent variables).
Int. Conf. on Probabilistic Graphical Models, Lugano, Sept. 6–9, 2016 27
Model fit to the data
Int. Conf. on Probabilistic Graphical Models, Lugano, Sept. 6–9, 2016 28
Model fit to the data
0 500 1000 1500 2000
−1.5
e+07
−1.0
e+07
−5.0
e+06
0.0e
+00
Time (seconds)
Glo
bal l
ower
bou
nd Alg. BS(data%)/LRSVI 1%/0.55SVI 1%/0.75SVI 1%/0.99SVI 5%/0.55SVI 5%/0.75SVI 5%/0.99SVI 10%/0.55SVI 10%/0.75SVI 10%/0.99d−VMP
I Int. Conf. on Probabilistic Graphical Models, Lugano, Sept. 6–9, 2016 29
Test marginal log-likelihood
BS (% data) LR Log-Likel.
SVI
1 %0.55 -180902.870.75 -298564.030.99 -426979.52
5 %0.55 -177302.240.75 -333264.160.99 -628105.70
10 %0.55 -347035.220.75 -397525.450.99 -538087.13
d-VMP 1.0 67265.34
Int. Conf. on Probabilistic Graphical Models, Lugano, Sept. 6–9, 2016 30
Mixtures of learnt posteriors for one attribute
Attribute range
Den
sity
−10000 0 10000 20000
0.00
000.
0004
0.00
080.
0012
●
●
●
●
SVI 1% SVI 5%SVI 10%VMP 1%
Int. Conf. on Probabilistic Graphical Models, Lugano, Sept. 6–9, 2016 31
Scalability settings
I Generated data set of 42 million samples per client and 12variables.
I “Unrolled” model of more than 1 billion (10
9) nodes
(75% latent variables).I AMIDST Toolbox with Apache Flink.I Amazon Web Services (AWS) as distributed computing
environment.
Int. Conf. on Probabilistic Graphical Models, Lugano, Sept. 6–9, 2016 32
Scalability results
Int. Conf. on Probabilistic Graphical Models, Lugano, Sept. 6–9, 2016 33
Outline
1 Motivation
2 Variational Message Passing
3 d-VMP
4 Experimental results
5 Conclusions
Int. Conf. on Probabilistic Graphical Models, Lugano, Sept. 6–9, 2016 34
Conclusions
I Variational methods can be scaled using distributedcomputation instead of sampling techniques.
I Bayesian learning in model with more than 1 billion nodes(75% of hidden).
Int. Conf. on Probabilistic Graphical Models, Lugano, Sept. 6–9, 2016 35
Thank you for your attention
Questions?
You can download our open source Java toolbox:amidsttoolbox.com
Acknowledgments: This project has received funding from the European Union’s
Seventh Framework Programme for research, technological development and
demonstration under grant agreement no 619209
Int. Conf. on Probabilistic Graphical Models, Lugano, Sept. 6–9, 2016 36