ff discusses grid cells and their role in spatial representation and movement integration within the brain. It explores their interaction with place cells, the mechanism of oscillatory interference for path integration, and more.
5 Likes
Hello and thank you for making this information available to the public and the manner in which your book, the TBP, is written. I am a layperson and don't wish to cloud the discussion, but here I go.
I will try to be as concise as I can. But, when every column has a voice, even if they choose not to use it, I look at the process in other areas and sometimes bring it in for scrutiny in another application. These many voices (columns) working in parallel, able to form group think (or not) is a form of robustness, not redundancy. When columns are in agreement I liken it to multiple conductors having their orchestra play the same piece, starting at the same time. There is also an efficiency within the process that accepts some errors, or inaccuracies, in certain areas as acceptable (or non action errors). It is perfectly acceptable, in our many orchestra example that one tuba player goes off paper in something that they play. In the grand scheme, perfection is not required and once they hear a particular marker in the music that should get a nice reset to be coordinated again.
Let me talk briefly about my understanding about models and efficiency and questions for applications into AI. I can’t help but feel that there needs to be a robust efficiency mechanism for not only object and understanding, but for flow. When you first brought up the modules, I had a strong sense of coordination. This may not be about precision, or correct mapping. Can it be a coordinated effort to be good enough in memory, sensory inputs, conceptual reasoning and predictive summary blended with the focus at hand/end game. A tuba player going rogue is fine, because it doesn’t really matter if you have inaccuracies along the way. Now, if they never “reset” when place markers are encountered…that may become a problem. Now let’s mix in scale with the conception of the end point (understanding that some people may prefer to break end game down into manageable chunks, while others can look at the whole picture). Scale makes things manageable to end point. What is detail in one scale is forgettable noise in the two steps larger, etc. Coordinated prediction doesn’t have to be accurate either, it will refine itself through feedback loops with the mentioned memory, sensory, conceptual evaluation. The module recalibrates continually allowing it to have a errors. This makes the system efficient and resilient not getting hung up on wrong information. Memory can be fuzzy, sensory data incomplete and predictions inaccurate, but the module is still effective. The degree and ratio of errors doesn’t even need consistency, only that flow and feedback moves toward the most probabilistic outcomes it’s capable of at the time.
Applying this process in a different scenario of recognizing a dog, car, coffee cup or stapler. If I need to apply the robust nature and resiliency discussed above to identification of object, concepts, function and application. Again, we don’t want to be needlessly accurate, especially at first. We want good enough, we want flow and efficiency in a resilient model that will blend pattern recognition, memory, application of function, sensory input and prediction. Abstract conceptualization of dog, coffee cup, car should highlight efficiency and resiliency. Being flexible and making generalizations allow a system to identify a flying car when seeing it for the first time. In this way early learning may look like baby steps, but the end result could provide a more flexible framework in a more efficient model. Initial exploration, iterative learning, feedback through interaction (get things wrong and refine), pattern recognition from this and self improvement/learning. Concept of what constitutes a car is much better than recognizing every car that was available from past to present.
That is a nutshell version of a few thoughts. Thank you so much for your time and allowing me to be stimulated. I will go back to finish the video and the 1000 Brain Project book.
~cheers~
1 Like
I love those kinds of videos!
“So trying to imagine now that cortex is this, Sheet of cells, we say it’s six layers, let’s pretend it’s six layers. when it gets, as I said, when it gets to the sort of the end of its sheet, it turns into, it becomes less structured, less uniform. And the first thing, one of the first things it becomes is, literally follow the sheet and it becomes the entorhinal cortex, there’s other things first, but, and one can argue that the entorhinal cortex is, Let’s say roughly three layers thick, and then it continues on and it folds over on top of itself, and the hippocampus is now on top of, it’s folded, like a sheet of paper, and now you have three layers of the hippocampus on top of the three layers of the entorhinal cortex. And this is a very rough approximation, but it’s true. And so the idea there is, one of the ideas that we didn’t come up with, but it seems like it’s a reasonable idea, is that the three layers of the hippocampus on top of the three layers of the entorhinal cortex, then some, at one point in past history, became the six layers of the cortex. And so there’s some analogies between the place cells in the hippocampus being the upper layers of the cortical column and the entorhinal cells and the entorhinal cortex being the lower layers, and I’ve always got that in the back of my mind. So not only did evolution discover this thing, but physically there’s a suggestion that, the cortex became the fusion of these two three layer structures that are aligned on top of one another.” (6’30’’)
Having the six-layered neocortex evolve from the combination of a three-layer entorhinal cortex-like structure and a three-layer hippocampus-like structure sounds like a nice story, but I’m afraid that the 3-layer entorhinal cortex is a wrong assumption. Even if there are significant differences between the mEC and the neocortex, the mEC is often considered to have 6 layers that roughly relate to the 6-layered neocortex (with a tiny L4 and an unclear separation between L5 and L6, in addition to varying cell types like many stellate cells instead of pyramidal cells in superficial layers). Or am I missing something?
Why not try to model the mEC directly as a collection of cortical columns, similar to neocortical areas, using the same framework? Could the entorhinal cortex, thanks to its expressive and meaningful grid cells, serve as the Rosetta stone for deciphering the neocortex?
My previous comment in the other thread is completely related to this video: 2023/01 - A Comprehensive Overview of Monty and the Evidence-Based Learning Module - #5 by mthiboust
1 Like