AlexNet was the winning model of the 2012 ImageNet competition. It has a deep convolutional neural network architecture consisting of five convolutional layers and three fully connected layers. AlexNet contains over 60 million parameters and has more filters and layers than previous networks, helping to address the problem of overfitting. It also employs techniques like dropout and data augmentation to improve performance.

1. ALEXNET
(WINNER OF 2012 IMAGENET COMPETITION
SUSHIL KULKARNI

2. THE ARCHITECTURE
• The AlexNet architecture was proposed by Alex Krizhevsky, Llya
Sutskever ,and Geoffrey E. Hilton and research paper written based on
this architecture is “imageNet Classification With Deep Convolution
Neural Networks”
• The architecture has:
1.Five Convolution Layers
2.Three Fully Connected Layers (Final Layer is softmax)

3. CONTD…
• There are 60 Million Parameters in AlexNet so it is called as Deep
Neural network
• It has more filters per layer as compared to LeNet. Convolution layers
are more in AlxNet
• Dropout feature is added to ensure overfitting problem
• Data Augmentation( Mirroring image, to increase the training volume)
is also done

4. THE ARCHITECTURE
Confusing right  , Yes it will be for beginners ! We shall make it
easier !

5. THE ARCHITECTURE
227*227*
3
Conv,
11*11,
96
filters,
stride
=
4
Max
pooling,
3*3,
stride
=
2
Conv,
5*5,
256
filters,
padding
=2
Max
pooling,
3*3,
stride
=
2
Conv,
3*3,
384
filters,
padding
=1
Conv,
3*3,
384
filters,
padding
=1
Conv,
3*3,
256
filters,
padding
=1
Max
pooling,
3*3,
stride
=
2
Fully
Connected
Layer_
1,
4096
Fully
Connected
Layer
_2,
4096
Output
Layer,
Softmax,1000

6. DIMENSIONS AND DETAILS
Layer Filters size Kernel Size Stride Activation
Input Image 1 227*227*3 -
1 Convolution 96 55*55*96 11*11 4 relu
Max Pooling 96 27*27*256 3*3 2
2 Convolution 256 27*27*256 5*5 1 relu
Max Pooling 256 13*13*256 3*3 2
3 Convolution 384 13*13*384 3*3 1 relu
4 Convolution 384 13*13*384 3*3 1 relu
5 Convolution 256 13*13*256 3*3 1 relu
Max Pooling 256 6*6*256 3*3 2
6 Fully
Connected
- 9216 (
Neurons)
- - relu
7 Fully
Connected
- 4096 (
Neurons)
- - relu
Out Put Fully
Connected
- 1000 (
Neurons)
- - softmax

7. DIMENSIONS AND DETAILS
(
𝑛+2𝑝−𝑓
𝑠
+ 1,
𝑛+2𝑝−𝑓
𝑠
+ 1 )
(
27+2(2)−5
1
+ 1,
27+2(2)−5
1
+ 1) = 27*27*256
Without Padding Formula
(
𝑛−𝑓
𝑠
+ 1,
𝑛−𝑓
𝑠
+ 1 )
(
227−11
4
+ 1,
227−11
4
+ 1) = 55*55*96
With Padding Formula

8. THE ARCHITECTURE
227*227*
3
Conv,
11*11,
96
filters,
stride
=
4
Max
pooling,
3*3,
stride
=
2
Conv,
5*5,
256
filters,
padding
=2
Max
pooling,
3*3,
stride
=
2
Conv,
3*3,
384
filters,
padding
=1
Conv,
3*3,
384
filters,
padding
=1
Conv,
3*3,
256
filters,
padding
=1
Max
pooling,
3*3,
stride
=
2
Fully
Connected
Layer_
1,
4096
Fully
Connected
Layer
_2,
4096
Output
Layer,
Softmax,1000
(
𝑛−𝑓
𝑠
+ 1,
𝑛−𝑓
𝑠
+ 1 )
= (
227−11
4
+ 1,
227−11
4
+ 1)
= 55*55*96
(
55−3
2
+ 1,
55−3
2
+ 1)
= 27*527*96
(
𝑛+2𝑝−𝑓
𝑠
+ 1,
𝑛+2𝑝−𝑓
𝑠
+ 1 )
=(
27+2(2)−5
1
+ 1,
27+2(2)−5
1
+ 1)
= 27*27*256
(
27−3
2
+ 1,
27−3
2
+ 1)
= 13*13*256
(
13+2(1)−3
1
+ 1,
13+2(1)−3
1
+ 1)
= 13*13*384
13*13*256
(
13−3
2
+ 1,
13−3
2
+ 1)
= 6*6*256
6*6*256=9216