@FFiebig and I were talking about Deep Predictive Learning: A Comprehensive Model of Three Visual Streams yesterday in the office, and he kept referring to this video as we discussed the pulivnar.
I am happy to see that you are finally coming up to speed on the 3VS paper. A lot going on on and the authors assume that you know all the background. I am looking forward to your excellent presentation making this available In a more approachable way.
This fits HTM into a larger and more useful framework. Also, it ties the thalamus into the picture and gives a clear narrative on how it interacts with the cortex.
One aspect that Dr Sherman brought up was the soup of neurochemical receptors. It could be thought of as making the system hopelessly complicated.
In my mind I have simplified it as follows, start with these considerations:
- The thalamus is the gating and distribution network. You can think of it as attention and automatic gain control if that works for you.
- It is not likely that these complicated signaling pathways are somehow directly wired into the learned engrams. The influence has to be indirect or tonic in some way.
- Even though these various signaling pathways are not directly wired into the engrams they must be able to call up the related engrams as attractor nodes when the subcortical structures call for attention to them later.
- The only thing that the cortex gets is what is passed to it through the thalamus. If some of this gating is “anded” with valance driven signaling (various neurochemical signals) then the cortex will learn these things better and see these things better when those gates are enabled.
Q: Where do these circulating chemical and specialized signaling axons come from?
A: subcortical structures that do command and control of the brain.
So - I think that the presence of signaling from subcortical structures enhances attention in both learning and recall.
- In the attention/learning part the local areas that are being boosted by these messengers have an advantage and tend to prioritize learning in the attached cortex. This is a very fine-grained localized effect. This loads the perceived memory with valence.
- Later - in attention to something needed. The subcortical structures release chemical messengers related to this need and perception of things with this preloaded emotional valence are prioritized with attention to enter into the global workspace. This emotional valence was learned along with the perception of some prior perception.
- These islands of valance flavored learning will be attractors in future recognition/recall. Further activation with chemical messengers gates these attractors.
This scheme allows the rather dumb subcortical structures to reach out and interact with the very complicated things that the cortex is learning; more of the dumb boss/smart advisor model. What is going on at the HC/EC is related but at the episode level rather than the learned object level.
I do see this. Hunger drives extremely complex survival behaviors, which are all stored in cortex. And the “dumb boss, smart advisor” metaphor is entirely appropriate.
I have to add that this is highly speculative as I have never seen anyone else propose this sort of thing anywhere. I came up with this as the only thing I could think of that fit everything I have read AND inspiration from the dumb boss/smart advisor model.
If you look at the brain though the dumb boss/smart advisor lens a lot of things fit into the narrative. It may turn out to be pixie dust but it does seem to have predictive power.
Regarding the links with the 3VS paper and the role of thalamus in routing the error signals, I am surprised that the role of matrix thalamic cells projecting to L1 is absent from the 3VS paper (maybe it is due to the fact that the paper is already two-year old and those findings are recent).
Because the apical dendritic tuft is probably involved in the credit assignment problem and because all thalamic nuclei have matrix cells projecting to L1 where apical dendritic tufts are located, we should take this into account in the task of routing errors.
Some related materials for this discussion:
Those two slides are amazing. I can’t wait to see the whole stack. Great job!
In your opinion, because of what you remarked, does it make the 3VS paper wrong in some places or just incomplete? I’m about to read it, which for me is a huge investment of focus.
Thanks for your comment and happy to see your enthusiasm!
The 3VS paper likely contains some errors. It is normal for this kind of speculative paper. But as @Bitking said in the other thread, it is probably one of the best paper of its kind, so I highly encourage you to invest some time in it. It will be useful for sure.
I read this paper several times in September and it was very enlightening because it puts diverse ideas together. It helped me a lot in my neuroscience self-training because it directed me to relevant topics I didn’t know.
Though, there were still many paragraphs that weren’t clear for me. I’ll need to reread it again because I learned new things since my first read. Maybe my ideas about the error routing through the matrix cells will get clearer after that!
I watched the Murray Sherman video today, thanks for the link. I was kind of floored when he said that hardly anybody is studying multi sensory integration!
Well, perhaps nobody he knows. He needs to get out more.