4
$\begingroup$

Are classic signal processing/statistics based approaches to optimum detection/estimation still relevant/important compared to ML based approaches using DL?

There was a time when speech processing was done using HMM. While not completely unused most speech processing now rests firmly in the DL area. I'm wondering if every area of classic signal processing will be taken over by these approaches over the classic feature finding historically used.

$\endgroup$
4
$\begingroup$

This is an open discussion.

In my opinion, any field we couldn't figure near optimal solution (And in most cases we only have Linear Optimal solution) using Deep Learning will replace classic methods given enough data to work with.

The power of data driven features vs. intuitive (Though mostly right yet never cover everything) will prevail.
I don't see it as a problem, Deep Learning isn't throw it all and it work machine, it still requires domain knowledge to effectively build the training process, the architecture and mostly the model to optimize. Which, if out think about it, just like in a classic method: Build test case, build a model, build a solver.

So one can see it as an evolutionary step where the optimization and the model are greatly improved yet the method has the same rational.

I believe near the hardware Signal Processing will stay as it is for longer time (Decimation, Filtering, etc...). Though in image processing DL is doing steps towards this as well. But in RF I think it will stay longer. But the next steps (For instance, Matched Filter) can be replaced (In Low SNR?) by DL.

$\endgroup$
3
  • $\begingroup$ I agree, it seems signal processing will work if you have a very good mathematical model describing the process that's sufficiently tractable. But even then I have seen DL being used in areas like CT reconstruction and radar so IDK anymore. I think communications is maybe one area that will mostly stay with classic DSP for awhile since the models are often simple enough to analytically find optimal solutions and recover properties. $\endgroup$ Oct 4 '20 at 6:55
  • $\begingroup$ Communication certainly will DL. For CT and RADAR we have the simple models. That's what I wrote, for most classic signal processing, we have simple models of the reality. For example, in RF we mostly neglect 2nd and 3rd order phenomenons. DL has the potential to take them into account. What will be left is only the Pre Processing near the hardware or places we can't gather data (Like One Shot). $\endgroup$
    – Royi
    Oct 4 '20 at 8:15
  • $\begingroup$ Good points. But I feel doing all signal processing problems as DL is a big step backwards. DL has many problems, it's limited to the quality of the dataset, it's a black box and I'm not sure about the reliability of these models. DL models are known to fail embarrassingly. DL, like AI is very over-hyped too. I don't believe AI and DL will change the world for the better. $\endgroup$ Jan 6 at 13:38
0
$\begingroup$

Yes, classic signal processing/statistics is still relevant. For example in optimum detection/estimation, if a sub task demands you to find out the nature of a group of signals obtained(by nature I mean the distribution- if it's Gaussian or Uniform or so on), then the preferred way of finding that out would be using measures like Mean, Variance, Kurtosis, Skewness. That saves lot of resources and computation which would have been spent had someone opted for training and predicting(or in short use a machine/deep learning algorithm). Hope that helps your question.

$\endgroup$

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.