WTA model used in core Nengo

General Discussion
1

In "A Spiking Independent Accumulator Model for Winner-Take-All Computation " by @jgosmann et al. two models of WTA memory are presented: LCA and IA. Are both of these models are implemented in nengo-spa 0.3? Is the LCA model equivalent to the base WTA memory? What’s the relation between the nengo-spa WTA memory and the one In core Nengo 2.5 implemented by the AssociativeMemory module.

I ask, because I’m 80% sure (from looking at the code) the answer to these questions is that LCA == nengo-spa WTA == nengo core Associative Memory, but I’d like to be sure before I cite the publication instead of spending a paragraph explaining how the WTA memory works.

I’m also sure I’ve asked this to @jgosmann at least twice, but I can’t seem to remember the answer, so I’m asking it in a public place so it’s easier to find.

2

The core Nengo WTA and Nengo SPA WTA should be the same. They are similar to the LCA, but not identical (might just be a matter of parameter choice, e.g. set threshold=0, account for synaptic time constants etc.).

1 Like
3

As @jgosmann pointed out during lunch today, the inhibition happens in different state spaces for the Nengo WTA and LCA, because the LCA inhibition does not act on the neurons directly, but instead acts on the state-space.

4

I just checked the code and that is actually not true. The inhibition is happening in the state space.