0
$\begingroup$

Intro:

Lots of Machine Learning methods are inspired Biology, Nature, Physics, Neurology...

I just thought of a Deep Learning approach inspired on religion:

Rebirth Network

Some eastern religions state that our soul has a deeper purpose and it's immortal. When we are born, we briefly forget this purpose so we can expand further our consciousness. (Something like this, I don't really know)

Anyway, I was thinking on a DNN based on this concept. For explicit comparison, the soul would be a deeper layer (very slow training due to Vanishing Gradient problem). The "shallow" layers would be our body, living some few experiences, learning until converge and than dying.

Hypothesis / Question:

  • If we reset the shallow layers periodically. Could that somehow lead to a better deep layer on the long run?
  • Could this approach be actually useful in any scenario?
  • If that's a bad idea, can you explain me why?

Consider any scenario where it could be useful, for example:

  • Run each "reincarnation" with a different dataset, or even topology
  • Adding / removing ResNet to the model
  • Setting the Reset condition based on training convergence
$\endgroup$
2
  • 1
    $\begingroup$ It is a bit of a random idea, without solid theory, just a guess. No-one will be able to answer this question based just on theory, and I doubt many people have tried it. Your best bet is to try the idea yourself on a well-known problem, such as an ImageNet challenge. You could use a SotA model, like ResNet50, and see if any of your proposed modifications improve upon it $\endgroup$ Aug 27 at 7:10
  • $\begingroup$ That certainly seems new. You will have to try it and see how well it works, and spend weeks or months playing with the parameters. At least I think that's how AI research is done :) $\endgroup$
    – user253751
    Aug 27 at 8:33

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.