0
$\begingroup$

I started to study NN recently. So I understand principles with which I should define input and output layers.

But I can't find any guide/directions how to build hidden layers: how many layers do I need, how many neurons per layer, what activation functions should I use etc for different types of tasks.

So I am searching for some guide like: try to start with ... layers of ... neurons. If you get ... result, please increase number of neurons ..., but if you get... result, you should try to decrease number of neurons. Or something similar.

$\endgroup$
  • $\begingroup$ Who is the ninja voter not commenting why the vote was given? On down vote constructive explanation is a rule not exception. $\endgroup$ – mico Dec 28 '17 at 21:27
  • $\begingroup$ I didn't downvote, but I think that this question is good and definitely doesn't deserve a downvote. Therefore +1 from me. $\endgroup$ – user10671 Dec 30 '17 at 2:02
1
$\begingroup$

In general no. The research area that you seem to be looking for is called generalization, which is very much so still an active area of research. Actual architecture design strongly depends on the dataset itself and available resources, and thus there isn't a general rule of thumb that works for every case.

$\endgroup$
0
$\begingroup$

Google (Scholar) can answer this kind of questions. For example, the inverted pendulum problem is normally solved with one neuron which has 4 weights, Getting started with OpenAI gym Other problems like image recognition needs more neurons, the exact number of hidden layers is given in the papers. If somebody finds a completely new architecture, which outperforms a previous solution he would also publish it. So the exact number of neurons, and the used neural network architecture can be discussed not on a mathematical base but in a literature review.

$\endgroup$

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.