Aside from obvious biological plausibility, from a computational standpoint, what's the motivation of using the Neural Engineering Framework (NEF) instead of Artificial Neural Networks (ANNs) for computing functions? From what I can tell, they both approximate functions, have the ability to learn new functions and the ability to create self-organizing networks. What makes the NEF different?


First, let's take a look at the basic principles of NEF.

The first two principles (Representation and Computation) do seem analogous to trained ANN models. Additionally, with the hPES learning rule that I've described here, they seem to have the same learning (gradient-descent) capabilities.

Where the NEF differentiates itself the most from ANNs is when it starts to describe dynamics (oscillators, attractors). For a description on how it does this and what it accomplishes, check out Terry Stewart's course notes found here.

Describing dynamics is important in the brain, since simple oscillators and integrators seem to be the foundation of many different cognitive abilities such as working memory and motor control.

  • $\begingroup$ What the NEF does is translate neural coding into the solving of differential equaions (well functions more generally) You might argue that it's a specific implementation of ANN. $\endgroup$ – Keegan Keplinger Sep 12 '14 at 15:51
  • $\begingroup$ @KeeganKeplinger When you say "specific implementation" of ANNs, it could be interpreted as saying that they both have the same functional capabilities, but I'm not sure I understand your evidence for this. Could you help me out? $\endgroup$ – Seanny123 Oct 15 '14 at 18:46
  • $\begingroup$ an ANN is just a network of simple neurons (usually designed in some connective pattern, designed to serve a purpose). NEF utilizes an ANN to solve differential equations. So of course, ANN, being the more general term, has more capabilities... but once you choose an instance of an ANN for a specific purpose, you begin to limit the capabilities of the ANN towards your purpose. $\endgroup$ – Keegan Keplinger Oct 15 '14 at 19:03
  • $\begingroup$ So when you say ANN, you don't mean specifically the back-prop type. You just mean any system that has nodes that are neurons that are then arranged in a manner to imitate neural computation? $\endgroup$ – Seanny123 Oct 15 '14 at 19:17
  • 1
    $\begingroup$ When I see ANN I think of individual neurons modeled for simplicity and problem solving, not physiology (so not ion current models based on actual channel kinetics). Usually the goal is to compute something, not understand neurons - I think that the core definition of ANN is in the goal of the network design, not the type of neurons used. The two I mostly see in such approaches are integrate-and-fire and firing rate models. Though I'm no authority here, my research in ion-conductance models. I know sometimes you can deduce group neuron behavior using ANN's, so there's likely a fuzzy line $\endgroup$ – Keegan Keplinger Oct 15 '14 at 19:28

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.