Fig 1: Input Image

AlexNet is a convolutional neural network (CNN), which was designed by Alex Krizhevsky. It got popular after the competition of the ImageNet Large Scale Visual Recognition Challenge on September 30, 2012.

AlexNet is an 8 layer deep network. It has-

• Five Convolution layers
• Three fully connected layers(last one is the Softmax Layer)

In the following table, all the 25 layers(including pooling layers, relu layers etc) with its description has been stated. 2,6,10,12 14 are the Conv Layer and 17,20,23 are the fully connected layer.

It is to be noted that the input images should be of 227x227 pixels.

In this article, we will try to visualize the outputs of the layers. It's not actually possible to show all the 25 layers output. So, I will show only the first few layers output here.

Conv Layer Activation Maps

First Conv activations

In the figure 2.1(a), first Conv activations are shown. Here, only the edges and colors are detected. The gray channels indicate that these channels have not been activated at all. The black pixels are mostly positively activated and the white ones are negatively activated.

Figure 2.1(a)

In the figure 2.1(b), most activated channel of this Conv layer is shown. from this picture, it can easily be said that, all the edges are clear.

Figure: 2.1(b)

Activation after the ReLU Layer (layer no 3)

Figure: 2.2

Activation after the Normalization Layer (layer no 4)

Figure 2.3

Activation after the Pooling Layer (layer no 5)

Fig 2.4

Activation of the Fully-Connected Layer(layer no 23)

In figure 2.5, some complex pattern are detected. from the picture we can assume that this layer has detected

1. Tench
2. Goldfish
3. Great white shark
4. Tiger shark
5. Hammerhead
6. Electric Ray

Fig: 2.5

Filter Patterns

In 1st Conv layer 96 filters have been used. In the figure 3, first 36 filters have been shown. These filters are blob edge detectors

Figure 3: Filters of 2nd Layer

Weights of 1st Conv Layer

The weights of 1st Conv layer is (11 x 11 x 3 x 96).from figure 5, we can assume that only vertical and horizontal lines are learned from this weight.

Figure 5: Weights of 1st Conv layer

Thanks for reading!

And don't forget to share your thoughts with me!!!