1
$\begingroup$

As part of a project I'm working on, I have built a "fully connected layer" (multilayered perceptron) network for image classification. Even though I know how to build an convolutional NN, for various reasons I had to use a fully connected one.

My question is: given the network, learned to classify with test accuracy of 86%, how can I visualize what each hidden neuron is looking for? And, more broadly, how can I understand better how the network decides how to classify? I have seen many ways to discover this in CNN, but not in fully connected networks.

I used tensorflow to build the network.

$\endgroup$
1
$\begingroup$

One of the problems with the fully connected networks is that except the output layer, individual units in the hidden layers don't have necessarily any semantic meaning themselves; it is rather the space they induce that is interesting. In other words, in the hidden layers, the network can represent useful things as a combination of activations of many units, without the individual units having a useful meaning without the rest.

However, there are some things you can do still (from :

Moreover, on this site, there is a nice live demo of activations of a simple FCN.

| cite | improve this answer | |
$\endgroup$

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.