I recently got into Neural networks. As much as I have understood, the learning process is based on the change in weights according to stimulus and algorithm used in learning. Does this in any way represent the real learning process of our brain?


Yes and no.

Neural networks that use edge weights simulate synaptic plasticity - a key mechanism in the brain operation. But it isn't the only one.

In practice, cognition is the combination of various mechanisms, some we are still unsure of.

Perhaps the most obvious missing ingredient of weighted networks is the lack of frequency/time domains similar to that affecting the brain. Neurons involve time-dependent operation - for instance, both the action potential and the following refractory period are time-bound. The brain also involves synchrony, e.g. the phase-locked discharges of a neuron set.

There're also issues like different axon lengths that play a part. And many more.

In addition, and possibly above all, the actual topography matters a great deal - how everything is connected.

More advance brain simulations involve many of these extra components - mimicking more accurately a real biological brain.

  • $\begingroup$ Recurrent Neural networks use delay operators to allow a gap before the feedback reaches. The neuron doesn't activate next time until the feedback. Isn't that quite close to the refractory period of a cell? $\endgroup$ Dec 2 '15 at 17:25
  • $\begingroup$ @kneelb4darth - Absolutely. My answer applies to neural networks of the most basic kind - those based on edge weights only. There are a multitude of factors you can incorporate to make it more brain-like. Delay operators are just one. $\endgroup$
    – Izhaki
    Dec 2 '15 at 17:33
  • $\begingroup$ Could you please expand a bit on the synchrony part. Googling it didn't get me much :) $\endgroup$ Dec 2 '15 at 17:41
  • $\begingroup$ @kneelb4darth Take a look at en.wikipedia.org/wiki/Neural_oscillation#Phase_resetting $\endgroup$
    – Justas
    Dec 2 '15 at 18:43
  • $\begingroup$ @kneelb4darth, this is an excellent answer on the topic: Importance of Neural Synchrony to Cognition $\endgroup$
    – Izhaki
    Dec 2 '15 at 19:48

The approach of the artificial neural networks that you describe and their application is called Connectionism. There are a number of cognitive architectures that have used this approach to explain cognition, such as Leabra.

The question about whether this is really what the brain is doing is another question entirely. Does the brain do back-propagation? This is currently up for debate. Does it contain structures such as LSTMs and other Recurrent Neural Networks? Probably not, as these aren't very efficient uses of memory and neurons. For a further discussion of biological plausibility, I would recommend reading "How to Build a Brain" by Chris Eliasmith.


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.