General Neural Networks Compositions of linear maps and - - PowerPoint PPT Presentation

general neural networks
SMART_READER_LITE
LIVE PREVIEW

General Neural Networks Compositions of linear maps and - - PowerPoint PPT Presentation

Pooling and Invariance in Convolutional Neural Networks General Neural Networks Compositions of linear maps and component-wise non- linearities Neural Networks Common Non-Linearities Rectifier Linear Unit Sigmoid Hyperbolic tangent


slide-1
SLIDE 1

Pooling and Invariance in Convolutional Neural Networks

slide-2
SLIDE 2

General Neural Networks

Compositions of linear maps and component-wise non- linearities

slide-3
SLIDE 3

Neural Networks

slide-4
SLIDE 4

Common Non-Linearities

Rectifier Linear Unit Sigmoid Hyperbolic tangent

slide-5
SLIDE 5

Biological Inspiration

Neurons diagram Rectified Linear Unit

slide-6
SLIDE 6

Sparse Connections

Not required to be fully connected

slide-7
SLIDE 7

Backpropagation

slide-8
SLIDE 8

Representations Learnt

slide-9
SLIDE 9

Representation Learning

Images far apart in Euclidean space Need to find representation such that members of same class are mapped to similar values

slide-10
SLIDE 10

Convolutional Neural Network

Compositions of: Convolutions by Linear Filter Thresholding Non-Linearities Spatial Pooling

slide-11
SLIDE 11

Convolution by Linear Filter

slide-12
SLIDE 12

Convolution by Linear Filter

slide-13
SLIDE 13

Example

slide-14
SLIDE 14

Convolutional Neural Networks

1) Convolution by Linear Filter 2) Apply non-linearity 3) Pooling

slide-15
SLIDE 15

Convolutional Neural Network

slide-16
SLIDE 16

Convolutional Neural Networks

slide-17
SLIDE 17

Imagenet Network

slide-18
SLIDE 18

Successes

Computer vision Speech Chemistry

slide-19
SLIDE 19

Object Classification

a

slide-20
SLIDE 20

Segmentation

slide-21
SLIDE 21

Object Detection

slide-22
SLIDE 22

Speech

slide-23
SLIDE 23

Physical Chemistry

Successfully predict atomization energy, polarizability, frontier orbital eigenvalues, ionization potential, electron affinity and excitation energies from molecular structure.

slide-24
SLIDE 24

Visualization of First Layer

slide-25
SLIDE 25

Standard Pooling Mechanisms

Ave pooling Max pooling

slide-26
SLIDE 26

Example

slide-27
SLIDE 27

Heterogenous Pooling

Some filters passed to Ave pooling Others filters passed to Max pooling

slide-28
SLIDE 28

Pooling Continuum

Accordingly, LeCun et al. 2012 ran experiment with variety of “p” values.

slide-29
SLIDE 29

Results along spectrum

Optimal for this SVHN dataset was p = 4.

slide-30
SLIDE 30

L_p learnt pooling

Why not learn optimal p for each filter map?

slide-31
SLIDE 31

Stochastic Pooling

slide-32
SLIDE 32

Stochastic Pooling

Expectation at Test Time

slide-33
SLIDE 33

Entropy Pooling

Extend to variable p In particular, Alternative:

slide-34
SLIDE 34

Max-Out pooling

Pooling across filters Substantial Improvement in performance and allowed depth

slide-35
SLIDE 35

Example

slide-36
SLIDE 36

Compete Neurons

Neurons can suppress other responses

slide-37
SLIDE 37

Visualizations of Filters

Early Layers

slide-38
SLIDE 38

Visualizations

slide-39
SLIDE 39

Visualizations

slide-40
SLIDE 40

Invariance under Rigid Motion

Goodfellow et al. 2009 demonstrated the CNN are invariant under Indeed, depth of NN critical to establishing such invariance

slide-41
SLIDE 41

Unstable under Deformation

Szegedy et al.

slide-42
SLIDE 42

Lipshitz Bounds for Layers

Max and ReLU are contractive FC Layers: usual linear

  • perator norm

Conv Layers: Parseval’s and DFT yield explicit formula

slide-43
SLIDE 43

Solutions?

Regularize Lipshitz operator

slide-44
SLIDE 44

Coding Symmetry

Convolutional Wavelet Networks

slide-45
SLIDE 45

Architecture

Wavelet convolutions composed with modulus

  • perator
slide-46
SLIDE 46

Gabor Wavelets

Trigonometric function in Gaussian Envelope

slide-47
SLIDE 47

Group Convolution

slide-48
SLIDE 48

Group Invariant Scattering

Main Result: Conv. Wavelet Networks are translation invariant functions in L_2(R^2) Furthermore, CWN can be made invariant to action under any compact lie group.

slide-49
SLIDE 49

Textures

Sifre and Mallat

slide-50
SLIDE 50

Basis for images

Learns similar representation as Imagenet CNN

slide-51
SLIDE 51

Learnt Invariances

slide-52
SLIDE 52

Optimal network

1) Encoded symmetry 2) Regularize Lipshitz coefficients 3) Compete Neurons in final layers

slide-53
SLIDE 53

Final words