Hello I want to design a AI.The neural network of my AI will consist of 1 input layer of neurons and 1 output layer.

What is very unique about the neural network is that the # of input neurons will change while the AI is learning.

For example suppose we 3 snapshots over a learning epoch:

I need to look for the algorithm which would decide when to add or delete a neuron.Any ideas?

  • $\begingroup$ Welcome to AI Stack Exchange. It is worth explaining why you want this flexible architecture, and the nature of problems you expect it to solve. There are existing learning methods that allow NN architecture to change, but I have not seen any that would change the input layer, since that is usually defined by the problem being solved - so it appears you want to change the data or problem, perhaps as a form of curriculum learning (where you start with a simple problem and use that to bootstrap solutions to more complex ones)? $\endgroup$ Commented Jun 27 at 8:15
  • $\begingroup$ @NeilSlater imagine a "dumb" AI first booted taking values through sensors and being asked to answer a question.The AI doesnt know which sensors to pay attention to so to start the process , it takes the values of only 1 sensor then it outputs some value.But if the correct output isnt a function of the values of the 1 selected sensor , then it needs to reevaluate its strategy and take into consideration values coming from a different subset from the values set of all sensors.Until the AI is rewarded ,it has to keep searching you need to be able to add/remove neurons from the input.Thanks. $\endgroup$ Commented Jun 27 at 9:21
  • $\begingroup$ So you will know what the ful llist of sensors is at any point, but you want the AI to discover the useful ones during learning? Does this have to involve starting with one sensor, or could it start equally by using them all and ruling out the ones that don't help with establishing correct answer? $\endgroup$ Commented Jun 27 at 13:54
  • $\begingroup$ @NeilSlater Yes I want the AI to discover the useful ones during learning $\endgroup$ Commented Jun 27 at 17:19
  • $\begingroup$ Is your requirement for this question to build an NN that performs that useful input discovery, (and the details of how it works secondary), or are you only interested in approaches that start with one sensor connected and build up from there, as per your drawings? $\endgroup$ Commented Jun 27 at 18:02


You must log in to answer this question.