2
$\begingroup$

Two practical questions regarding the use of autoencoders. For the size of the NN, let's say I have 50 inputs, so layer sizes go like 50-25-12-6-2-6-12-25-50

  1. how much data I generally need to train it and how does it depend on the size of net?

  2. I'd like to weigh samples differently - e.g. there is a time component to the problem the last samples are most recent, so should weigh more. Is there an easy way to incorporate it?

$\endgroup$
1
$\begingroup$

Each network has a limit that it can learn. To check that limit, you need to try. Especially those kind of simple networks, it can be achieved so quickly. Idea is that, you need to give X data and check the accuracy and increase it to 2X data and check the accuracy. If there is no change in accuracy, it means that your network has reached its limit and you may consider making it deeper and wider.

By nature of the training, the last sample that you bring to training is the result that optimizer will try to converge. Hence, optimizers are already doing what you are searching for. However, if you want to create some weights for them, try considering RNNs. Their nature is suitable for that purpose.

$\endgroup$

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.