Size of training data for neural network
I have a time series forecasting task. I have heard that LSTM neural networks are one of the best algorithms to predict time series. But how much data I need to train the neural network well?
Answers
Hi @orisKon
Mmh, I think the more data you have the better. But it depends on the task complexity. If your data has vivid patterns that are easy to detect and predict, probably you can use less training data. The advice is to try and compare the results with the results given by others (machine learning or statistical algorithms)
You can also try to use transfer learning for improving the quality of your network's predictions when you have not enough data.
How I can perform transfer learning correctly in my case?
A general rule is to use as similar data as possible. Ideally, it should have the same time-dependent patterns.
Also, you can make several layers of the transfer learning. Each next layer should be closer to your own dataset. The scarcity of training data for time series forecasting is a common issue, so I think you can find a lot of useful information on the web.
Thank you all!
Are you shure you want to close this post?
Are you shure you want to pin this post?
Report comment
Thank you for reporting this comment, the Imaginghub Community Management will verify your request.
Log in with Google