LENET ARCHITECTURE

INTRODUCTION TO LENET

Lenet 5 was introduced in the year 1998 by Yann Lecun. Lenet is the one the oldest pretrained convolutional neural networks and state of the art of Deep learning algorithms. Lenet architecture has 5 layers, hence it is named as Lenet5.The architecture consists of convolutional layers, maxpooling layers and fully connected layers. It uses different filters and kernels to extract features from the image.

ARCHITECTURE

The input image is of dimensions 32*32*1. Here 1 represents the grace scale image, hence the number of channels in the input image is 1.Lenet architecture consists of three convolutional layers, 2 average maxpooling layers and 2 fully connected layers. The last fully connected layer acts as the output layer.It uses 3 different filters namely 6, 16, 120 to extract features from the image. It uses 2 different kernel sizes namely 2*2 & 5*5. The Kernel size 5*5 is used for convolutional layers while 2*2 is used in average maxpooling layer.It uses stride 1 for convolutional layer & stride 2 for average maxpooling layer.Lenet architecture has no padding and it uses tanh activation function for convolutional and average maxpooling layer and soft max activation function for the output layer.

 F-filters ,K=kernels ,s=stride

                                                    LENET ARCHITECTURE

DEEP DIVE INTO LAYERS OF LENET

                             The first layer in lenet comprises of 6 filters  with the kernel size  of 5x5 to extract different features by applying different filters from the image with tanh activation function of stride 1.The resulting output image dimensions is 28x28x6.here 6 represents the six different extracted features from the image.The output image dimensions is calculated by

                                                                    n-f+2p\s+1

here,

n-input image dimensions

f-filter size

p-padding

s-stride

for first layer it is calculated as,

                                               = 32-5+2(0)/1+1=28

     n=32, f=5 , p=0, s=1

The input to the second layer is 28x28x6.Here we are applying average maxpooling of kernel size 2x2 with stride 2  to reduce the dimensions of the image.The output size of the image is 14x14x6.the input size  given to the third layer is 14x14x6. The third layer consists of 16 filters with kernel size of 5x5 with tanh activation function of stride 1.The output of the third layer is of dimensions 10x10x16.The calculation is as follows,

                                               =14-5+2(0)/1+1=10

Here,n=14,f=5,p=0,s=1

The input to the fourth layer is of size 10x10x16 .Then we  are applying average maxpooling of kernel size 2x2 with stride 2  to reduce the dimensions of the image.the output size of the image is 5x5x16.The input to the fifth layer is of size 5x5x16.Again,we are applying 120 filters with kernel size of 5x5 with tanh activation function of stride. The output size  of the fifth layer is 1x1x120.the sixth layer is a fully connected layer with 84 neurons .Finally there is a ouput layer with  softmax activation function of 10 ouputs.