Bitking, thanks for this, it is giving me tons to think about.
Thanks for the link!
Different paper but the same general information.
A definition of abstract is “disassociated from any specific instance” (source: Merriam Webster). Numenta’s 2017 Columns & Layers paper explains one form of abstractions: view-point invariance. Viewpoint invariance is abstraction because there is a single abstract representation which responds to many concrete sensory representations of an object. And this does happen in the cortex, in layers 2/3.
My hypothesis is that the basal ganglia drives the frontal cortex. The basal ganglia is known to perform reinforcement learning, meaning that it predicts rewards and penalties. One of the outputs of the basal ganglia is decisions, and the basal ganglia is capable of turning motor areas on/off to enact those decisions. The Striatum is the first piece of the Basal Ganglia; it receives input from the entire cortex in a topologically precise manner and this is the only pathway from the cortex into the basal ganglia. Therefore all information which you use to make a decision needs to pass through the Striatum. The Striatum helps the basal ganglia by identifying relevant information and filtering out irrelevant info. I hypothesize that the Striatum projects to the frontal cortex via thalamic relay cells. I don’t have a lot of evidence but I think that this hypothesis explains a lot. It explains how the frontal cortex can use unsupervised learning and also understand rewards. Under this hypothesis, the frontal cortex only sees things which you’ve identified as either rewarding or punishing.
The reason I say that the magic step of abstraction happens outside of the neocortex is because the Striatum is responsible for seeing attributes such as danger, attractiveness, and reliability. The frontal-cortex is responsible for building a model of these important attributes, and that model includes grid cells which encode locations in this “emotional space”. An example of a location which frontal-cortical-grid-cells might encode is the reliability of a tool, which degrades as the tool is used. The frontal-cortex identifies when the tool is likely to break; and the Striatum sees the frontal-cortical model and evaluates when its time to stop working and replace the tool (before it breaks).
Probably you are right there (I can’t fully understand it). Nevertheless, if we take a look back at HM case (Henry Molaison), the 2/3 of the HC/EC/ complex was removed (https://en.wikipedia.org/wiki/Henry_Molaison).
Being a bit naïve here, my take is that EC and grid cells are mostly involved in new episodic memory formation. HM kept the capacity to form procedural memory. I think that hippocampal replays use the grid cells to determine where in the cortex the replay should be done (and where the sensory stream should be fast-stored in the HC). Therefore, is just a translation mechanism from HC addresses to cortex addresses (some sort of TLB ) You might observe grid alike behavior elsewhere because the “TLB” can have multiple levels (?)
I don’t think that HM lost the capacity of using abstract ideas. Only the ability to create new episodic memories. Replays are needed because up in the hierarchy the events are really sparse in time. Note that replays even occur when we are awake.
Just my 2 cents
That’s not completely true if I recall correctly. The cortex also projects to the subthalamic nucleus. It’s not part of the basal ganglia but the cortex also projects to zona incerta, which inhibits the thalamus or can disinhibit by cells inhibiting other cells in ZI. These are just nitpicks intended to help expand ideas.
The striatum indirectly (via other parts of the basal ganglia) projects to thalamus. A small number (one?) of thalamic nuclei are thought to lack any proximal excitatory inputs, meaning they are driven by the basal ganglia (or rather, driven right when inhibition from the basal ganglia turns off because of tonic/burst mode).
The prefrontal cortex has some regions involved in identifying different types of value (like intrinsic value like food or value not determined beforehand like food brand preferences). I don’t think all PFC regions exclusively encode things with positive or negative value, but there are at least some regions with strong bias towards responding to those things.
In what you describe, I see the model building part as the abstraction step and the value-detecting step as a gate. Maybe that gate is what makes the cortex do abstraction rather than just encoding what it sees, but I don’t see how that works. I think of abstraction as things like recognizing two completely different objects as the same category (like a cup you’ve never seen before or whether something is food) or making inferences, like understanding the idea conveyed by a sentence or figuring out where an object might be. I think these things require representations of predictions and making predictions from predictions, probably imagination. Two objects are in the same category if they share predictions (which is basically context except it doesn’t have to occur every time it is represented).
Well, the best way to check a hypothesis is to build a working model
At the end of the day, from the computational perspective, it’s not so important where, but what and how it’s happening. If you believe you understand how it works, it’s a good idea to implement it in code.
Regarding ideas and navigation, a new paper co-authored by Moser (Nobel Prize for his work on grid cells) has just been published in Science.
“Researchers find evidence that our experiences and knowledge are oganized in the brain in a spatial fashion”
Probably worth reading, but behind a paywall…
I’m sure that the brain uses spatial coding for abstract patterns. My two points in this respect are as follows:
- The grid cells aren’t the foundation of this organization and play supporting role only.
- There isn’t a need in the six-fold symmetry of grid cells in computer modelling to get the same functionality.
And thank you for the link!
BTW, the first time I heard that it’s not about the grid sells in the first place from Edvard Moser - I thought if he was thinking about it in this direction, it couldn’t be a biased point of view
Not paywalled article: https://www.cell.com/neuron/fulltext/S0896-6273(18)30856-0
I have posted something on this topic before.
Perhaps this and that add some explanation to the paper you linked. I don’t think that it is that crazy or random.
“The hippocampus/EC structures are tuned for one-shot episodic memory; the cortex is tuned for slower Hebbian learning.”
“What is the episodic part of our neural mechanism supposed to make of our mental furniture being shifted around and modified while these little bits of new learning is being consolidated? Perceiving these recalled patterns as they are being recalled and reinforced is what we call “dreaming.”
Feel free to post comments on the linked thread or here, based on the content of the post.
My point was that place/grid cells are just “artifacts” of hippocampal replay (i.e. are only involved in learning). They have little to no effect on “prediction/inference”.
I don’t think that anyone here will argue that this in not true. The HC/EC is situated in the correct relationship with the temporal lobe and the outputs of the WHAT & WHERE streams as a processor of digested high level representations. A good chunk of this is spatial information.
The same basic computational structures are distributed throughout the cortex and used in different way based on the representation of data at that level of processing.
Let me shortcircuit the question about what I mean by levels - I mean the distance or number of maps between the map in question and the primary sensory modality that is being processed.
Speaking of “processing,” I feel that there is a big concept lurking here. I have been pondering what is being done in the cortex for a long time. I think that there is something like Shannon’s seminal paper on the theory of what information is being conveyed and the amount of transformation that is done in each map waiting to be discovered.
… just isn’t aligned at all with current Numenta ideas.
Don’t think so. The replay are done across all the regions in the cortex (such as motor cortex), not only “high” levels in the hierarchy.
I don’t know how SDR idea fits in Shannon Theory … way above my level
I like to think of the HC/EC as a key like you would see in database theory. Sure - the tables hold the data but the keys organize it.
If you have been reading my stuff on hex-grid coding you know that I am not entirely alligned with Numenta either.
They have a lot of good stuff to offer. When they start getting some community stuff up to play with I expect some significant changes to make it useful.
Perhaps, we need to show tangible results … the problem is that is not so easy
FYI: Yet another article describing some recent research results relating our spatial reasoning/navigating abilities to our capacity for abstract thoughts. http://nautil.us/blog/new-evidence-for-the-geometry-of-thought
A point to consider regarding the Hippocampus and Entorhinal cortex: It is common to attribute spatial processing to this area. The discovery of place and grid cells has biased many to think that this is where this sort of representation is processed.
It is tempting to think that the cognitive map resides there.
I suggest that you take a more nuanced view of the relationship of these areas and spatial processing.
Patient HM continued to have the cognitive ability to process spatial concepts even though he had lost the ability to form new memories. You would think that the loss of the “cognitive map processor” or “spatial processor” would cripple these capabilities if this were true.
We should note that H.M. did particularly well on such timed tasks as the Block-Design subtest of the Wechsler Scale, or Milner’s version of Hebb’s triangular-blocks task (HEBB , p. 278 ; MILNER ). His success on these tests clearly indicates that spatial relationships as such cause him little trouble.
See some of the details here:
I offer that this has been an area where this representation has been observed but the rest of the cortex is also processing the same information. As I have posted elsewhere in this forum (here and here), I think that the HC/EC complex is important for short term (~one day) storage of this information. I also think that this is where experience interacts with emotional coloring from the amygdala to form valuation of objects and events.
The general spatial processing task is performed in the cortex, likely distributed over the temporal and parietal lobes.