2026/05 - Open Theory Questions

For this brainstorming week, @jhawkins talked about open theory questions. They discussed topics such as shared learning, voting with relative pose, attention, and basic modifications to our columns or LM models.

Summary Video

Main Video

0:00 Introduction
0:28 Open Issues We Need to Address
2:41 Shared Learning
4:12 Voting with Relative Pose
7:12 Integration of HCC Fast Learning with Cortical Slow Learning
12:25 Attention
15:29 Classes
18:47 Basic Modifications to our Column/LM model
21:53 Other Open Questions
30:36 Penn State University Project on Extracting Richer Features
37:36 Spatial Pooler Properties
42:55 Object Skeleton and Behavior Models
44:19 Object Skeleton Morphology Discussion
51:35 Reframing Behavior Generalization Problem as Object Model Representation Problem
55:01 Simplifying Morphology Comparison Using Compact Representation of Skeleton
58:40 Further Discussion
1:06:50 We Can Infer a Lot from Very Minimal Object Representations

2 Likes

I value how TBP sometimes pulls back from the details to ask lurking questions in hopes of seeing new, overarching patterns. The video somehow reminded me of the work of psychologist Lawrence Barsalou. I thought some of his work might enlighten the debate by pulling two sides together.

My overall take from the discussion was that it became an interesting debate as to how the brain generalizes behaviors across wildly different objects.

On one side, there’s a push for abstract representation — a morphology model stripped of specific physical features, leaving basically just oriented edges at a location.

On the other side, there’s resistance to this kind of abstraction, that the brain relies heavily on specific physical realities and not merely generic models.

From my (less-than-expert) read of psychologist Lawrence Barsalou’s theory of Grounded Cognition, both are correct. Here’s how:

I think that Barsalou resolves this dichotomy of views by explaining that human concepts (which he calls simulators) inherently contain both levels of structure:

  • an underlying abstract framework and
  • specific, physical instantiations.

Consider cognitive scientist Benjamin Bergen’s studies on how people simulate imaginary objects like “Pigasus” the flying pig. Barsalou argues we do this by dynamically constructing a simulation on the fly by taking our grounded memory of a pig and attaching the grounded memory of bird wings to it. Some might even add other embellishments of grounded memory such as a cape, goggles, and such. Each such grounded memory has an abstraction associated with it. I think of it as a highly lossy compression — one of a pig, one of a wings, one of a cape, etc.

So, to create the simulation of Pigasus, we take these lossy abstractions and stitch them together and then apply the phenomenal part of our memory of the physical thing to it. Like adding skin to a skeleton.

Forgive me if I’m misrepresenting Barsalou or Bergen in this. But I mainly wanted to share a couple of references to Barsalou, if it would help in the team’s thinking about these matters.


Barsalou, L. W. (2026). Grounded Cognition. In M. C. Frank & A. Majid (Eds.), Open Encyclopedia of Cognitive Science. MIT Press. https://doi.org/10.21428/e2759450.e89203d9

Barsalou, L. W. (1999). Perceptual symbol systems. Behavioral and Brain Sciences, 22(4), 577–660. http://ruccs.rutgers.edu/images/personal-zenon-pylyshyn/class-info/FP2012/FP2012_readings/Barsalou_BBS1999.pdf

Bergen, B. K. (2012). Louder than words: The new science of how the mind makes meaning. New York, NY: Basic Books.

1 Like