SLIDE 1
- Global Features
- The perceptron model
- Neural Networks – multilayer perceptron model (MLP)
- Backpropagation
Neural Networks + Convolutional Neural Networks Last Class Global - - PowerPoint PPT Presentation
Neural Networks + Convolutional Neural Networks Last Class Global - - PowerPoint PPT Presentation
CS4501: Introduction to Computer Vision Neural Networks + Convolutional Neural Networks Last Class Global Features The perceptron model Neural Networks multilayer perceptron model (MLP) Backpropagation Todays Class
SLIDE 2
SLIDE 3
- Neural Networks – multilayer perceptron model (MLP)
- Backpropagation
- Convolutional Neural Networks
Today’s Class
SLIDE 4
Perceptron Model
Frank Rosenblatt (1957) - Cornell University More: https://en.wikipedia.org/wiki/Perceptron
! " = $1, if * +,",
- ,./
+ 1 > 0 0,
- therwise
"; "< "= ">
*
+; +< += +> Activation function
SLIDE 5
Perceptron Model
Frank Rosenblatt (1957) - Cornell University More: https://en.wikipedia.org/wiki/Perceptron
! " = $1, if * +,",
- ,./
+ 1 > 0 0,
- therwise
"; "< "= ">
*
+; +< += +>
!?
SLIDE 6
Perceptron Model
Frank Rosenblatt (1957) - Cornell University More: https://en.wikipedia.org/wiki/Perceptron
! " = $1, if * +,",
- ,./
+ 1 > 0 0,
- therwise
"; "< "= ">
*
+; +< += +> Activation function
SLIDE 7
Activation Functions
ReLU(x) = max(0, x) Tanh(x) Sigmoid(x) Step(x)
SLIDE 8
Pytorch - Perceptron
SLIDE 9
Two-layer Multi-layer Perceptron (MLP)
!" !# !$ !%
&
'" '# '$ '%
& & & &
( )" ”hidden" layer (" Loss / Criterion
SLIDE 10
Forward pass
!" !# !$ !%
&
'" '# '$ '%
& & & &
( )" ("
*+ = &
- "+.'+
/ +01
+ 3" !+ = 567896:(*+) =" = &
- #+!+
/ +01
+ 3# (" = 567896:(=+) >9?? = >((", ( )")
SLIDE 11
Backward pass
!" !# !$ !%
&
'" '# '$ '%
& & & &
( )" ("
*+ *,- = * *,- /012304(,-) *+ *!7
89 8:;
=
8 8:; /012304(<-) 89 8= ); 89 8= ); = 8 8= ); +((", (
)") *+ *'7 = ( * *'7 & ?"-7'-
@
- AB
+ D7) *+ *,- *+ *?"-E = *'7 *?"-E *+ *'7 *+ *!7 = ( * *!7 & ?#-!-
@
- AB
+ D#) *+ *<" *+ *?#- = *!7 *?#- *+ *!7
GradInputs GradParams
SLIDE 12
Pytorch – Two-layer MLP + Regression
SLIDE 13
Pytorch – Two-layer MLP + LogSoftmax
# of Hidden Units
SLIDE 14
Pytorch – Two-layer MLP + LogSoftmax
LogSoftmax + Negative Likelihood Loss
SLIDE 15
Bookmark Opportunity!
SLIDE 16
SGD training code (Project 4)
SLIDE 17
Convolutional (Neural) Networks
SLIDE 18
Convolutional Layer
SLIDE 19
Convolutional Layer
SLIDE 20
Convolutional Layer
Weights
SLIDE 21
Convolutional Layer
4 Weights
SLIDE 22
Convolutional Layer
4 1 Weights
SLIDE 23
Convolutional Layer (with 4 filters)
Input: 1x224x224 Output: 4x224x224 if zero padding, and stride = 1 weights: 4x1x9x9
SLIDE 24
Convolutional Layer (with 4 filters)
Input: 1x224x224 Output: 4x112x112 if zero padding, but stride = 2 weights: 4x1x9x9
SLIDE 25
Convolutional Layer in Torch
nInputPlane (e.g. 3 for RGB inputs) nOutputPlane (equals the number of convolutional filters for this layer)
nOutputPlane x
nInputPlane
kW kH
Input Output
SLIDE 26
nInputPlane (e.g. 3 for RGB inputs) nOutputPlane (equals the number of convolutional filters for this layer)
nOutputPlane x
nInputPlane
kW kH
Input Output
Convolution2D(nOutputPlane, kW, kH, input_shape = (3, 224, 224), subsample = 2, border_mode = valid)
Convolutional Layer in Keras
SLIDE 27
Convolutional Layer in pytorch
in_channels (e.g. 3 for RGB inputs)
- ut_channels (equals the number of
convolutional filters for this layer)
- ut_channels x
in_channels
kernel_size kernel_size
Input Output
SLIDE 28
Convolutional Network: LeNet
SLIDE 29
LeNet in Pytorch
SLIDE 30
SpatialMaxPooling Layer
take the max in this neighborhood 8 8 8 8 8
SLIDE 31
Convolutional Layers as Matrix Multiplication
https://petewarden.com/2015/04/20/why-gemm-is-at-the-heart-of-deep-learning/
SLIDE 32
Convolutional Layers as Matrix Multiplication
https://petewarden.com/2015/04/20/why-gemm-is-at-the-heart-of-deep-learning/
SLIDE 33
Convolutional Layers as Matrix Multiplication
https://petewarden.com/2015/04/20/why-gemm-is-at-the-heart-of-deep-learning/
Pros? Cons?
SLIDE 34
Questions?
34