My question is very general and it does not originate from a specific problem. Let's assume that, through experience, we have learned that some statistical property of a set of data is important in predicting some behavior in a system. For example, for time series data d1, d2, d3, ..., dn, we heuristically know that the average of the last n steps denoted by avg(d,n) and the standard deviation of the last m denoted by std(d,m) are significant in prediction. Now my questions are:

  1. Should a machine learning system, let's say LSTM, or reinforcement learning agent, be fed the raw data or data with other statistical properties? I am asking this because, if the statistical derivatives are useful in training then there is no limit on how many statistical properties we can define and feed to the training process.

  2. Do machine learning, again let's say LSTM, automatically learn about underlying statistics from just pure raw data?

  3. How do we deal with different data of different scales and dimensions, for example, simple average is in the same scale as the raw data but standard deviation is of different scale and dimension and so on so forth?

I appreciate your comments.

  • 1
    $\begingroup$ Usually the point is that the system learns to approximate whatever calculations it needs. $\endgroup$ Commented Sep 14, 2021 at 16:28
  • 1
    $\begingroup$ Hello, asking multiple questions in a single post is not recommended. Please try to split the questions into multiple posts. $\endgroup$
    – hanugm
    Commented Sep 14, 2021 at 21:57


You must log in to answer this question.