Lecture 3:
CNN: Back-propagation
boris. ginzburg@intel.com
1
Agenda
Introduction to gradient-based learning for Convolutional NN
Backpropagation for basic layers
– Softmax
– Fully Connected layer
– Pooling
– ReLU
– Convolutional layer
Implementation of back-propagation for Convolutional layer
CIFAR-10 training
2
Good Links
1. http://yann.lecun.com/exdb/publis/pdf/lecun-01a.pdf
2. http://www.iro.umontreal.ca/~pift6266/H10/notes/gradien
t.html#flowgraph
3
Gradient based training
Conv. NN is a function y = f(x0,w), where
x0 is image [28,28],
w – network parameters (weights, bias)
y – softmax output= probability that x belongs to one of 10 classes 0..9
4
Gradient based training
We want to find parameters W, to minimize an error
E (f(x0,w),y0) = -log (f(x0,w)- y0).
For this we will do iterative gradient descent:
−𝜕𝐸
w(t) = w(t-1) – λ * (t)
𝜕𝑤
How do we compute gradient of E wrt weights?
Loss function E is chain of functions. Let’ s go layer by layer,
from last layer back, and use the chain rule for gradient of
complex functions:
𝜕𝐸 𝜕𝐸 𝜕𝑦𝑙 (𝑤,𝑦𝑙−1 )
= ×
𝜕𝑦𝑙−1 𝜕𝑦𝑙 𝜕𝑦𝑙−1
𝜕𝐸 𝜕𝐸 𝜕𝑦𝑙 (𝑤,𝑦𝑙−1 )
= ×
𝜕𝑤𝑙 𝜕𝑦𝑙 𝜕𝑤𝑙
5
LeNet topology
Soft Max + LogLoss
Inner Product
ReLUP
BACKWARD
Inner Product
FORWARD
Pooling [2x2, stride 2]
Convolutional layer [5x5]
Pooling [2x2, stride 2]
Convolutional layer [5x5]
Data Layer
6
Layer:: Backward( )
class Layer {
Setup (bottom, top); // initialize layer
Forward (bottom, top); //compute : 𝑦𝑙 = 𝑓 𝑤𝑙 , 𝑦𝑙−1
Backward( top, bottom); //compute gradient
}
𝜕𝐸
Backward: we start from gradient from last layer and
𝜕𝑦𝑙
𝜕𝐸 𝜕𝐸
1) propagate gradient back : →
𝜕𝑦𝑙 𝜕𝑦𝑙−1
𝜕𝐸
2) compute the gradient of E wrt weights wl:
𝜕𝑤𝑙
7
Softmax with LogLoss Layer
Consider the last layer, softmax with log-loss (MNIST example ):
𝑒 𝑦0 9 𝑦𝑘
𝐸 = − log 𝑝𝑦0 = −log( 9 𝑒 𝑦𝑘 ) = −𝑦0 + log( 0𝑒 )
0
For all k=0..9 , except k0 (right answer) we want to decrease pk:
𝜕𝐸 𝑒 𝑦𝑘
= 9 𝑒 𝑦𝑘 = 𝑝𝑘 ,
𝜕 𝑦𝑘 0
for k=k0 (right answer) we want to increase pk:
𝜕𝐸
= −(1 − 𝑝𝑘0 )
𝜕 𝑦𝑘0
See http://ufldl.stanford.edu/wiki/index.php/Softmax_Regression
8
Inner product (Fully Connected) Layer
Fully connected layer is just Matrix – Vector multiplication:
𝑦𝑙 = 𝑊𝑙 ∗ 𝑦𝑙−1
𝜕𝐸 𝜕𝐸
So = ∗ 𝑊𝑙 𝑇
𝜕𝑦𝑙−1 𝜕𝑦𝑙
𝜕𝐸 𝜕𝐸
and = ∗ 𝑦𝑙−1
𝜕𝑊𝑙 𝜕𝑦𝑙
Note: we need 𝑦𝑙−1 , so we should keep them on forward pass.
9
ReLU Layer
Rectified Linear Unit :
𝑦𝑙 = max (0, 𝑦𝑙−1 )
= 0, 𝑖𝑓 (𝑦𝑙 < 0)
𝜕𝐿
so = 𝜕𝐿
𝜕𝑦𝑙−1 = , 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒
𝜕𝑦𝑙
10
Max-Pooling Layer
Forward :
for (p = 0; p< k; p++)
for (q = 0; q< k; q++)
yn (x, y) = max( yn (x, y), yn-1(x + p, y + q));
Backward:
𝜕𝐿
= 0, 𝑖𝑓 ( 𝑦𝑛 𝑥, 𝑦 ! = 𝑦𝑛−1 𝑥 + 𝑝, 𝑦 + 𝑞 )
(𝑥 + 𝑝, 𝑦 + 𝑞) = 𝜕𝐿
𝜕𝑦𝑛−1 = (𝑥, 𝑦), 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒
𝜕 𝑦𝑛
Quiz:
1. What will be gradient for Sum-pooling?
2. What will be gradient if pooling areas overlap? (e.g. stride =1)?
11
Convolutional Layer :: Backward
for (n = 0; n < N; n ++)
for (m = 0; m < M; m ++) M N
for(y = 0; y<Y; y++) K
W
for(x = 0; x<X; x++) X
for (p = 0; p< K; p++) Y
for (q = 0; q< K; q++)
yL (n; x, y) += yL-1(m, x+p, y+q) * w (n ,m; p, q);
Let’ s use the chain rule for convolutional layer
𝜕𝐸 𝜕𝐸 𝜕𝑦𝑙 (𝑤,𝑦𝑙−1 )
𝜕𝑦𝑙−1
=
𝜕𝑦𝑙
×
𝜕𝑦𝑙−1
;
𝜕𝐸 𝜕𝐸 𝜕𝑦𝑙 (𝑤, 𝑦𝑙−1 )
= ×
𝜕𝑤𝑙 𝜕𝑦𝑙 𝜕𝑤𝑙−1
12
Convolutional Layer :: Backward
Example: M=1, N=2, K=2.
Take one pixel in level (n-1). Which pixels in next level are influenced by it?
2x2
2x2
13
Convolutional Layer :: Backward
Let’ s use the chain rule for convolutional layer:
𝜕𝐸 𝜕𝐸
Gradient is sum of convolution with gradients over all
𝜕𝑦𝑙−1 𝜕𝑦𝑙
feature maps from “upper” layer:
𝜕𝐸 𝜕𝐸 𝜕𝑦𝑙 (𝑤,𝑦𝑙−1 ) 𝑁 𝜕𝐸
= × = 𝑛=1 𝑏𝑎𝑐𝑘_𝑐𝑜𝑟𝑟(𝑊, )
𝜕𝑦𝑙−1 𝜕𝑦𝑙 𝜕𝑦𝑙−1 𝜕𝑦 𝑙
Gradient of E wrt w is sum over all “pixels” (x,y) in the input
map :
𝜕𝐸 𝜕𝐸 𝜕𝑦𝑙 (𝑤,𝑦𝑙−1 ) 𝜕𝐸
= × = 0≤𝑥≤𝑋 𝑥, 𝑦 °𝑦𝑙−1 (𝑥, 𝑦)
𝜕𝑤𝑙 𝜕𝑙 𝜕𝑤𝑙 𝜕𝑦𝑙
0<𝑦<𝑌
14
Convolutional Layer :: Backward
How this is implemented:
backward ( ) { …
// im2col data to col_data
im2col_cpu( bottom_data , CHANNELS_, HEIGHT_, WIDTH_, KSIZE_, PAD_, STRIDE_,
col_data);
// gradient w.r.t. weight.:
caffe_cpu_gemm (CblasNoTrans, CblasTrans, M_, K_, N_, 1., top_diff, col_data , 1.,
weight_diff );
// gradient w.r.t. bottom data:
caffe_cpu_gemm (CblasTrans, CblasNoTrans, K_, N_, M_, 1., weight , top_diff , 0.,
col_diff );
// col2im back to the data
col2im_cpu(col_diff, CHANNELS_, HEIGHT_, WIDTH_, KSIZE_, PAD_, STRIDE_,
bottom_diff );
} 15
Convolutional Layer : im2col
Implementation is based on reduction of convolution layer to
matrix – matrix multiply ( See Chellapilla et all , “High Performance
Convolutional Neural Networks for Document Processing” )
16
CIFAR-10 Training
http://www.cs.toronto.edu/~kriz/cifar.html
https://www.kaggle.com/c/cifar-10
60000 32x32 colour images in 10 classes, with 6000 images per class.
There are:
50000 training images
10000 test images.
17
Exercises
1. Look at definition of following layers (Backward)
– sigmoid, tanh
2. Implement a new layer:
– softplus 𝑦𝑙 = log( 1 + 𝑒 𝑦𝑙−1 )
3. Train CIFAR-10 with different topologies
Project:
1. Port CIFAR-100 to caffe
18
