Would descriptors of the last hidden layer of two different CNN be the same?

Question

I am given a dataset of 2D medical images. I am asked to extract image descriptors from the hidden layer of the neural network pre-trained on the ImageNet dataset.
I consider to use two networks: ResNet-50 with imageNet accuracy of X and VGG-16 with imageNet accuracy of Y (X < Y). I wonder if, for any image, descriptors from the last hidden layer of ResNet-50 are the same as the descriptors from the last hidden layer of VGG-16.
I come from an NLP background, don't hesitate to explain to me as if I was 5 years old :p

Simon Larsson · Answer

No, there is nothing that says hidden layers should match even if they are used for the same task. And in the example you gave, ResNet-50 and VGG-16, they don't match. VGG-16 uses a fully connected layer with 4096 units. You can see it here:

While ResNet-50 uses global average pooling which produces a 1-dimensional vector with 2048 units. You can see it here:

Would descriptors of the last hidden layer of two different CNN be the same?

One Answer

Add your own answers!

Ask a Question