Learning Learning Models from Models from Data Data Presenter: Dale Schuurmans Presenter: Dale Schuurmans
Learning Learning Models from Models from
DataDataPresenter: Dale SchuurmansPresenter: Dale Schuurmans
2
Vision StatementVision Statement
Helping the world understand data and make informed decisions
Core technology research:• Underlies prediction and control,• Underlies most applications, • Involves almost every PI
3
MotivationMotivationAlmost all data is now digital
Text, images, audio, video (media) Scientific, financial, commercial, industrial data Robotics (sensors, telemetry), surveillance
Data is complex Video, text, microarrays, …
Data is accumulating
Data analysis is a major global industry Web search, finance Social and economic impact
But task embodies hard scientific problems
4
Discovering intrinsic structure in dataClasses
Manifolds
Components
Constraints
Dynamics
(Usually unsupervised)
Learning a ModelLearning a Model
5
A General ThrustA General Thrust
Unifying challenge for AICMLUnderlies many other thrusts, applications
Fundamental researchNew principlesNew algorithmsNew ideas
AICML advantageCritical massOutstanding resourcesDiversity of applications
6
Projects and StatusProjects and Status
1. Large margin clustering (completed; led to Project 2)
2. Convex, discriminative EM (ongoing; poster #16)
3. Bicluster coding (ongoing; poster #8)
4. Manifold dimension estimation (ongoing; poster #13)
5. Action respecting embedding (ongoing; poster #14)
7
Projects and StatusProjects and Status
6. Temporal difference networks (ongoing; poster #19)
7. Opponent modeling (ongoing; posters #9, #11)
8. WWW visualization (ongoing; posters #3, #4)
And other projects …
8
AICML Personnel (cumulative)AICML Personnel (cumulative)
Primary PI’s Goebel, Greiner, Holte, Bowling,
Schuurmans, Sutton, Szepesvari5 PDFs17 Grad students4 Technical staff
9
ResourcesResources
Grants$225K CFI New Opportunities Grant$100K MITACS Grant$50K Google GrantSmall part of $5.5M Alberta Transplant InstituteSmall part of $1M Polyomix Grant
Facilities68 processor, 2TB, Opteron cluster 54 processor, dual core, 1.5 TB, Opteron cluster
10
Partners/CollaboratorsPartners/Collaborators
Partners:Alberta Transplant InstituteCross Cancer InstituteUniversity of WaterlooUniversity of MichiganUniversity of MunichNational ICT AustraliaGoogle
11
Early Highlights Early Highlights
Bicluster codingDramatic improvement to cancer patient
prognosis from microarrays:62% 86% SVM accuracy
Action respecting embeddingMajor impact in mobile robotics community
“ARE is the most interesting use of dimensionality reduction I have seen’’ – leading expert
13
The ChallengeThe ChallengeConstruct a map from experience (SLAM), but …
Can this be done entirely automatically using machine learning principles?
16
The ChallengeThe Challenge
What makes a good internal representation?
Experience:
Encoding:
Good map:1. Encoding has correct (low) dimensionality2. Actions have simple representation in encoding
17
Basic Dimensionality ReductionBasic Dimensionality Reduction
Maximum Variance Unfolding(Saul, Weinberger, 2005)
Maximize spread of dataPreserve local distances
Solves 1 but not 2
18
ApproachApproach
Also need actions to have a simple representationIdea: add distance preserving constraints
Distance preserving implies actions are linear in representation
23
The FutureThe Future
ScalingSemidefinite programming is slowCan solve trajectories of length ~100
Continuous actionsObstacles or traversal costsState aliasing