A Framework For Cortical Computation Poster - Questions

Science Numenta Theory
1

Hello Gentlemen,

congratulations to another exciting poster, (this one)! Thank you again for sharing the information, I really appreciate what you’re doing and that you’re sharing all this!

After having a look I was thinking about what you wrote and I have a couple of questions. If you could shed some more light on this topic. In the poster there is a sketch of the cortical column, all my questions are related to it
image

1, What’s in this sketch - just to get a confirmation if I understand this correctly. Let’s say information from the eye is going through thalamus and arrives to L4 (onto proximal dendrites). L4 is also getting grid-cell-like information from L6a (grid cell layer attending to the position of the sensed feature). The information arrives to L4 onto distal dendrites. The L4 cells then activate L3 cells where it is pooled (temporal pooling) and fed further to L5a onto proximal dendrites. L5a is also getting information from L6b which is a displacement-like layer but the information is related to the object, not just the feature and it’s arriving to distal dendrites.
Then, information from L3 and L5a is fed the next level up; while information from L5a travels via collaterals into different regions (like motor region + older brain structures).
So basically, L4 has sensed features preactivated by the position info from L6a. Features are pooled into objects in L3 and that is going to L5a where it’s pooled based on the displacements of various parts of the object. So the next region up in the hierarchy is receiving the information of what object I’m seeing (from L3) and in which state (L5a).
So far so good or did I make a mistake somewhere?

2, What do you currently think is happening in L5b? I’m just curious because I don’t see any information coming in or going out of this layer.

3, I can see the information is fed from L3 to L2. I know these two layers are difficult to distinguish (at least in living tissue), and usually they are refered together as L23. But is there any reason why you’ve separated L23 into two separate L2 and L3 layers? If yes, is the L2 information used anywhere?

4, The green arrow between L3 and L5a in both ways - Shouldn’t there be a green arrow going down from L3 into L5a and a blue arrow going from L5a to L3? I.e. info from L3 arrives to prox dendrites in L5a, but L5a is going back to distal dendrites (and not proximals) in L3?

Like I said, thank you for your contribution, I think this is pretty cool, but I’d just like to check if my understanding is the same as to what you had in mind.

Thank you in advance!

MH

5 Likes
2

MH, great questions. This figure illustrates major documented pathways in the cortex. We are trying to understand all these connections and layers but we still have a ways to go. It was included in the poster as “this is what the cortex looks like everywhere, we want to understand this structure”. We think we understand some of this pretty well but other parts not.

  • L3 projects to L5a (These layer 5 cells are called thick tufted cells or cortical-thalamic-cortical CTC cells. BTW L5a and L5b are reversed in primates and rodents so better to use L5ctc or L5tt). We think L3 is an SDR representing the observed and learned object. We think L5tt is a set of displacements of the components of the object represented in L3. L3 to L5tt is a one-to-many projection.

  • L5b or L5 cortical-cortical or L5ccc cells. We don’t know what these are. I have theory that I have written about elsewhere in this forum that these cells are used for timing.

  • L3 vs. L2. Some people distinguish L3a vs. L3b. L3b gets feed forward input like L4, and L3a doesn’t. We don’t know what the distinction is. My current guess is that L3 is what we perceive, that is, stable objects at a particular orientation in the world. I say this is what we perceive because L3 is considered by many to be the main feed forward pathway. One guess for L2 is that it temporal pools over L3. L2 might represent an object independent of orientation. L2 is known to project broadly and across the callosum. If the left and right hands are touching the same object, they could vote on the object’s identity (L2), but not on the object at a particular orientation (L3) since the orientation would be different for the left and right hand. L2 would also be the better layer to classify, say “name this object”.

Yes! Great observation, you are following and understanding this deeply. I didn’t draw the arrow this way because it made the figure complicated. I experimented with drawing two arrows, one blue and one green, and with drawing one arrow with different colored ends. They both looked complicated and for the purposes of the poster I wasn’t trying to make this level of distinction.
Jeff

4 Likes