Learning Invariances and Hierarchies Pierre Baldi University of California, Irvine.

Post on 16-Dec-2015

213 views 0 download

Transcript of Learning Invariances and Hierarchies Pierre Baldi University of California, Irvine.

Learning Invariances and Hierarchies

Pierre BaldiUniversity of California, Irvine

Two Questions

1. “If we solve computer vision, we have pretty much solved AI.”

2. A-NNs vs B-NNs and Deep Learning.

If we solve computer vision…

If we solve computer vision…

• If we solve computer audition,….

If we solve computer vision…

• If we solve computer audition,….

• If we solve computer olfaction,…

If we solve computer vision…

• If we solve computer audition,….

• If we solve computer olfaction,…

• If we solve computer vision, how can we build computers that can prove Fermat’s last theorem?

Invariances

• Invariances in audition. We can recognize a tune invariantly with respect to: intensity, speed, tonality, harmonization, instrumentation, style, background.

• Invariances in olfaction. We can recognize an odor invariantly with respect to: concentrations, humidity, pressure, winds, mixtures, background.

Non-Invariances

• Invariances evolution did not care about (although we are still evolving!...)

– We cannot recognize faces upside down.– We cannot recognize tunes played in reverse.– We cannot recognize stereoisomers as such.

Enantiomers smell differently.

A-NNs vs B-NNs

Origin of Invariances• Weight sharing and translational invariance.• Can we quantify approximate weight sharing?• Can we use approximate weight sharing to improve

performance?• Some of the invariance comes • from the architecture. • Some may come from the • learning rules.

Learning Invariances

EHebbsymmetric connections

wij=wji

111

11-1

1-11

Acyclic orientation of the Hypercube O(H)

Isometry

Isometry

HebbHebb

O(H)

H

I(O(H))

I(H)

Deep Learning ≈ Deep Targets

Training set: (xi,yi) or i=1, . . ., m

?

Deep Target Algorithms

Deep Target Algorithms

Deep Target Algorithms

Deep Target Algorithms

Deep Target Algorithms

• In spite of the vanishing gradient problem, (and the Newton problem) nothing seems to beat back-propagation.

• Is backpropagation biologically plausible?

Mathematics of Dropout (Cheap Approximation to Training Full Ensemble)

Two Questions

1. “If we solve computer vision, we have pretty much solved AI.”

2. A-NNs vs B-NNs and Deep Learning.