Exotic Inhibitory Structures in Neocortex and Computational Implications

Well, maybe it's "exotic" only to neuroscientists. Inhibitory structure in neocortex has usually been seen as fairly homogenous and simple, where the wide variety of inhibitory interneuron types was viewed as misleading: at bottom, they all perform a simple regulatory function of keeping only a few representations active at once. Thus, in contrast to subcortical regions, inhibitory networks in cortex were thought to involve mostly "vanilla" axon-to-dendrite connections in a semi-regular latticework of connections. However, in the current issue of Nature Neuroscience, Conners & Cruikshank review recent evidence that neocortex might have some more complex inhibitory microcircuitry than previously thought.

The authors review evidence indicating that excitatory neurons might synapse directly on the the axons of inhibitory interneurons, "hijacking" them to cause the release of the inhibitory neurotransmitter GABA while bypassing the interneuron's dendrites and cell body. This kind of arrangement could provide "unusually fast, reliable and strong" inhibition directly from an excitatory neuron, and has not been previously observed.

The evidence used to support this surprising claim is rather compelling. Recordings from the visual cortex of mice show that nearly 28% of action potentials in pyramidal (excitatory) cells actually provoked a subsequent inhibitory postsynaptic current, suggesting a "axo-axo-somatic" synaptic triad architecture. This phenomenon was abolished by glutamate antagonists and kainate antagonists but remained to some extent under AMPA antagonists, suggesting that this is a different architecture from the more conventional excitation of inhibitory interneurons by pyramidal cells, which primarily takes place through AMPA receptors.

The original study also found immunocytochemical evidence to support this unusual arrangement of neurons in neocortex. But Conners & Cruikshank emphasize that that electron microscopy will be necessary to provide a clear picture of the structure of this new form of cortical microcircuitry. They also urge caution in that no such arrangement has ever been observed before now, and it is conceivable that these structures are specific to mice or visual cortex (or the combination).

Most interesting, however, are the computational implication of this discovery, as noted by Conners & Cruikshank. They suggest that the direct excitation of inhibitory interneuron axons might even cause a backwards-travelling wave of excitation (known as antidromic activity) in which other neurons which synapse onto the dendrites of the interneuron might themselves be inhibited. The potentially dazzling complexity allowed by this arrangement seems to share some similarities with the k-winners-take-all (kWTA) mechanism commonly implemented in neural network models, whereby the dominant neurons in a particular network will somehow enforce inhibition on the others. As I was recently reminded, kWTA is merely a "shorthand" for these more complex neurochemical processes, and it seems possible that a variety of mechanisms might ultimately underlie that function. Yet this heterogeneity in inhibitory structure also suggests that kWTA might be "missing the mark" - perhaps there is more structural organization than the "regular latticework" view of cortical inhibition would suggest.

This discovery highlights one aspect of computational neuroscience with which I am growing increasingly uncomfortable. The multiplicity of neurotransmitters, neuromodulators, and morphological structures in the brain makes it difficult to isolate those which are most critical for a particular computational operation. To use the current example, the function of keeping only a small number of representations active at once (conventionally modeled with kWTA) might be accomplished through a more conventional inhibitory interneuron latticework, through these newly hypoethesized axo-axo-somatic connections, through antidromic propagation, through some other as yet undiscovered mechanism, or through a combination of all of them. The computational neuroscientist is thus faced with an enormous array of candidate mechanisms to use in the modeling of such a simple process, which necessarily distances us from the larger (and in my view, more important) computational questions: how does mind emerge? How do we (or at least some of us) plan for the future? Where does priming end and conscious access begin? These kinds of questions are of course rooted in basic neurobiology, but that "basic" neurobiology is being steadily and incrementally revealed as far less basic than even highly complex computational theories had supposed.

More like this

These kinds of questions are of course rooted in basic neurobiology, but that "basic" neurobiology is being steadily and incrementally revealed as far less basic than even highly complex computational theories had supposed.

Personally, Chris, I love the complexity! Call it "job security" for neuroscientists and neuro-professionals.

Also, call it a challenge to the Jeff Hawkins' of the world, who want to waltz into consciousness research and create a machine intelligence almost as an afterthought. I like Hawkins' approach, but he and all the other machine intelligence optimists are going to have to face the complexity sooner or later.