Is there some guideline to designing architecture for neural networks?

I want to use LSTM network for predicting in time series. I have a small dictionary variation (8 values) but lot of their combinations.

How do I know how many memory block assemblies, how many memory cells each, input gates, forget gates, output gates and so on?

I am totally newbie, never used LSTM. I am in the very first phase - only research so far.

  • $\begingroup$ It depends on what kind of timeseries you're working with, and what the pattern length is between different outcomes. But I don't get what you are predicting, what do you mean with "I have a small dictionary variation"? $\endgroup$ – Thomas W May 7 '17 at 17:04

Unfortunately there are no solid guidelines for RNNs that would work every time. Try different combinations and see what works the best after a couple epochs and then train it more. Most often it depends on the pattern so it is mostly impossible to get the right network right away


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.