A thought about grid cells in low sensory cortex


Sorry if any of this is redundant with current thinking at Numenta and in the community. I’m intentionally not caught up on the new research and the related detailed theory posts because I want to minimize overlap between my false assumptions, things I chose to work on, etc. and those of others until I have a theory of layer 5 ready to refine and combine. That might take a while because I’m going to try to publish a review article as part of my research process. I shared my google docs notes in the past. If I haven’t done that yet for my current notes (the third notes doc), it would take two seconds to share a link if anyone wants but it might be easier for you to wait until I write a review or at least some summaries.

Brief Description
Once you understand it, this idea will be simple enough, but I’m not sure how to convey it clearly. Consider the sequence in which different points on the fingertip make contact with a given feature or object as you poke it. Feature shape + positioning relative to the skin -> a sequence of neuron firing onsets as their receptive fields on the skin make contact. The first neurons to begin firing inhibit those that would otherwise start firing soon after. Since those neurons are inhibited, they do not inhibit neurons, so there is a sort of periodicity. This periodicity is temporal, but because the sequence is based on locations, it is also spatial periodicity. If the first population of neurons to fire changes slightly because shapes/locations change slightly, neurons for the slightly different location instead activate first so the active neurons are completely different. There are naturally variations in periodicity because this whole system is pretty messy and depends on which neurons disynaptically inhibit which other neurons. So there are grid cell-like responses and we don’t really need anything special as a minimum framework.

How This Works And Evidence In Ramble Form Because I Haven’t Thought About It Long Enough To Explain It Well And Want To Write My Thoughts While They’re Fresh

Mostly About Evidence

Lately, I’ve been thinking that layer 5 (CC-S and TT at least) does not process sensory input details. Instead, it seems more involved in processing stimulus shape or mapping functions. High level allocentric regions have a map of objects on the cortical sheet (although I haven’t researched that yet so I could be wrong). Therefore, there needs to be some sort of coordinate transform from a sensor-centric map on the cortical sheet to an allocentric one. Layer 5 seems like the best fit for this function since it produces motor output and drives higher order thalamus. Some mysteries of the barrel cortex make more sense if layer 5 does this (specifically the weak responses of POm and L5 slender tufted cells except in certain circumstances.)

Assuming the cortex uses the described mechanism to form grid-like responses, the best evidence that it is in layer 5 is that they burst at stimulus onset and then fire less frequently. I have some issues with the idea of bursting in general (the extreme rising slope of current injection, some findings that even cells that burst in response to somatic injection at threshold only fire single spikes with insufficient slope of ramping injection, temperature and anesthetic-dependency, the fact that it is rare in normal functioning, findings by some slice studies and not by others that the apical tuft also increases the frequency of or directly causes regular spiking below a threshold, there are <100 hz frequency “pseudo-bursts” with less intense stimulation is some figures, and the general extremeness of the spiking and of the probably calcium-activated long afterhyperpolarization), but it’s still evidence and I don’t know enough about how neurons work to say otherwise.

More conceptually, Jeff Hawkins wrote a while ago that he thinks layer 5 represents stimulus onsets, if I remember correctly and understood, which is what suggested to me that role of low threshold bursting followed by regular spiking. So the cortex might treat stimulus onset differently from its continuation, perhaps explaining the duplicate input/output layer pairs. This idea also seems like it could contribute to precise timing, attention, behavior, and the link between behavior and perception. The fact that the same layer 5 cells project to motor structures and up the hierarchy via the thalamus seems like an important and useful mystery to me.

There is some evidence based on how it explains the difference between the strengths of V1 and barrel cortex primary thalamus input, but I want to keep the rambling to a minimum in this post. I’m not going to talk about layer 5 much more but that’s the only layer I know much about, so all of this is based on L5.

Mostly About Explaining It Conceptually

Sensory input can be divided into two types: onset and continuation. Continuation sensory input is more obvious. When you touch a surface, the spatial pattern of bumps and such while you hold your fingertip still is what I call continuation sensory input. However, this is not all of the available sensory information. The spatiotemporal pattern of contact is also very useful, based on an article I talked about in my previous post. Which part of your fingertip makes contact first (or second or last) says a lot about the shape of the feature.

You can think of the onset sensory input as a sequence of activity onsets on the scale of the cortical sheet map. This works in vision, hearing, and whisker contact, not just touch with skin. In hearing, people usually think of everything as a sequence so I don’t need to argue that the sequence evoked by sensory stimulation is important for hearing. All of these mechanisms to produce a sequence of activity from what we usually think of a single sensory input probably operate at the timescale of tens of milliseconds, which is nice for neurons.

Assuming V1 L5 surround responses are to barrel cortex L5, if the feature’s edge is further from the center of a cell’s receptive field, it has a longer latency response. This is a similar effect to the latencies induced by feature shape in touch, and it is clearly present in the barrel cortex when a whisker is deflected by an external stimulus. During whisker motion, this mechanism might exaggerate the effect too much because there is a whisker contact order like in skin contact, but there are a few mechanisms in the whisking system which may counterbalance this effect and even form the same mapping of head-centric sensory space whether whisker or not.


I understand the need for you to isolate from outside influences while you refine your thoughts… but maybe there’s indeed something of an overlap between yours and others, right there

Part above is somehow reminding me of Calvin… I guess @Bitking will be able to come up with pertinent comments on the matter.


This post is based on ideas of others (Bitking’s, Numenta’s, and a neuroscience article I wrote about in another topic I started, and probably other people’s). I still scan some things, just not everything.

That part is based on Bitking’s post HTM Columns into Grids! and its surround inhibition to generate grids, except with slightly temporally delayed inhibition and the inhibitory connections don’t need to be strict surround inhibition of neurons for parts of the sensor up to a distance equal to the grid cell spacing.