Can neural networks modify their own weights without back-propagation and gradient descent?

  • 1
    $\begingroup$ by "modifying weights" do you mean modifying the architecture? Second, backprop and gd do modify the weights-- thats legit all they do $\endgroup$ – mshlis Jul 25 '19 at 12:44
  • $\begingroup$ 1. I asked the question because I didn't know. 2. Then can the neural network modify weight? instead of backbrop, gd $\endgroup$ – Dimer Jul 25 '19 at 13:23
  • $\begingroup$ If your asking 1) can you learn a model architecture, then i can direct you to some cool papers or 2) if you can train a nn using other methods than bp, then i can refer you to genetic/evolutionary algorithms that might be of interest to you, but i still have no idea what your referring ro $\endgroup$ – mshlis Jul 25 '19 at 13:27
  • $\begingroup$ @mshlis Bp,gd is an algorithm, which is, can't we use neural networks instead of these algorithms to control weights? $\endgroup$ – Dimer Jul 25 '19 at 13:34
  • $\begingroup$ in a generative scheme? are you asking? (to clarify that is not at all what that quote you overheard was referring to) $\endgroup$ – mshlis Jul 25 '19 at 13:42

Artificial neural networks (ANNs) do not modify their own weights! Humans create algorithms that modify the weights or architecture of ANNs.

Having said that, you can change the weights of ANNs using other methods other than gradient descent combined with back-propagation, such as Hebbian learning, evolutionary algorithms or reinforcement learning.

| improve this answer | |

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.