ResearchComputing with neuronal cliques

What is the right abstraction for biological neural networks? While much of the machine learning community is concentrating on feedforward architectures, there is a rich set of feedback connections as well. What can be solved in networks with lateral connections within an area and feedforward/feedback connections between areas? Can optimization and inference problems be solved with only a few spikes?

We discovered that it is possible to solve a family of these problems with only a few spikes in 10s of msecs, provided the neurons are organized into (self-excitatory) cliques:

It follows from the theoretical observation that the network equations are related to polymatrix games:

...which in turn build upon relaxation labeling, our early mechanism to enforce consistency (note that "Hopfield Energy" is called "Average Local Consistency"):

Compatibility relationships are embodied in neuronal connections, and these can be developed directly from geometric considerations:

For a more recent review, see also: