2
$\begingroup$

In a recent answer I read on Stack Exchange, I read about a possible way to understand more clearly what happens in each hidden layer of a neural network.

Here's the excerpt-

You should watch what makes neuron activated in each layer depending on the input. As you know each neuron will be activated (once the DNN is trained) for specific input combinations. By visualizing that you can get an idea about what exactly each layer has learned in term of high-low level features.

Source - High-level features of a neural network

I wanted to know if there are any papers who have tried doing this (links would be really helpful). Meanwhile are there any other ways to understand what happens in each hidden layers?

$\endgroup$
  • $\begingroup$ if you are working on images there is a python code for visualizing each layers input for each applied filter, I think this can be helpful. otherwise I think it's hard to exactly understand what each layer learns especially when you have a little knowledge about image processing. $\endgroup$ – Hunar A.Ahmed Jun 8 '19 at 8:06
0
$\begingroup$

I wanted to know if there are any papers who have tried doing this (links would be really helpful).

The best known paper on this topic is Visualizing and Understanding Convolutional Networks by Matthew Zeiler and Rob Fergus. But you can find many more papers and online articles that do the same.

Meanwhile are there any other ways to understand what happens in each hidden layers?

Unfortunately, deep neural networks are like black boxes; we have little understanding of their inner workings. Visualizing the activations of the layers is probably the best way to tell what features they extract.

| improve this answer | |
$\endgroup$
  • $\begingroup$ Thanks, I will definitely check out that paper. $\endgroup$ – Franklin Varghese Jun 11 '19 at 11:32

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.