Flattening
1
3 0
-1 1 3
-1
3 1
0 3 Flattened
1
Fully Connected Feedforward
0 network
3
227*227*3=1,54,587
9216*6912=63700992=63.7*10^6
1 1
2 0
3
1 -1 -1 Filter 1 3 0
-1 1 -1 4 0
-1 -1 1 7 0
1 0 0 0 0 1 8 1
0 1 0 0 1 0 9 0
10 0 -1
0 0 1 1 0 0
1 0 0 0 1 0 13 0
0 1 0 0 1 0 14 0
0 0 1 0 1 0 15 1
6 x 6 image
16 1 Shared weights
36 0 5
Type of layers
Convolution Pooling Fully connected
Layer Layer Layer
2D 3D max Average
6
Why Pooling
Subsampling pixels will not change the object, however reduces parameter
count
bird Reduces the dimension of
feature maps
bird Translation invariant
Convergence rate of max
Subsampling pooling is faster compared to
average pooling, thus mostly
used.
8
Max Pooling
Input size : 8×8
Pooling window : 4×4
Stride : 2
Output : 4×4
9
The Whole CNN
Dog Cat
Convolution
Max Pooling Can repeat many
Fully Connected Feedforward times
network
Convolution
Max Pooling
Flattened
10
Convolutional Neural Network Architecture
All 3 RGB channels
Both fully connected layers
First, resized to 256 x Output to class
96 filters 256 filters 384 filters contain 512 neurons followed
256, then cropped to labels (age /
size 3x7x7 size 96x5x5 size 256x3x3 by ReLU and dropout layer
227 x 227 gender)
11
Machine Learning Deep Learning
•Enables machines to take decisions
•Enables machines to take decisions on
with the help of artificial neural
their own, based on past data
networks
•It needs only a small amount of data
•It needs a large amount of training
for training
data
•Works well on the low-end system, so
•Needs high-end machines because
you don't need large machines
it requires a lot of computing
•Most features need to be identified in
power
advance and manually coded
•The machine learns the features
•The problem is divided into two parts
from the data it is provided
and solved individually and then
•The problem is solved in an end-
combined
to-end manner