I've been working with multilayer perceptron (MLP) for a while. Whenever I assumed that the past values of a feature might be useful for predicting the future values of y, I would just create a new column in my data frame with Feature(t-1). This process would be repeated for further lags t-2,t-3...t-n.

Besides the obvious problem of the curse of dimensionality, I am worried that the MLP doesn't know how to weight those time lagged features that are now in a new separate column.

So in a nutshell:

  1. Is the above approach wrong?

  2. How does an RNN solve this?


1 Answer 1


Let's assume you have 3 features + considering 5 timestamps for each feature. That means you have 15 features as X values.

Machine learning techniques other than RNN can utilize the features as individual columns i.e. 15 columns.

RNN can use these data as 5 arrays containing 3 feature in them. We can feed this 5 array in RNN cell at each time step for a single record. The way RNN cell works when we feed second-time stamp data or subsequent time stamp data it gets processed important information from last time stamps.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.