Cognitive Architecture and HTM

The cognitive architecture approach is focused on building a “standard model” of mind. I think this is a place where HTM could be integrated towards building AGI. Specifically, modules of the architecture that are analogous to cortical processing could utilize the HTM approach. In my opinion, HTM can’t get anywhere close to AGI without being integrated into a complex cognitive architecture. It may also be that models could be greatly improved by integrating HTM theory and technology.

Here is a good overview of cognitive architecture.



Note: just for clarification, originally, CMC was named the “Standard Model of Mind” but later (after a community consensus) it got the CMC name. CMC underlying hypothesis supports that cognitive architectures provide the appropriate computational abstraction for defining a CMC, although the CMC is not itself such an architecture. A good place to start with Cognitive Architectures are these two papers:

  1. A Standard Model of the Mind: Toward a Common Computational Framework across Artificial Intelligence, Cognitive Science, Neuroscience, and Robotics
  2. 40 years of Cognitive Architectures: Core Cognitive Abilities and Practical Applications

Recently, I found myself thinking about a similar topic: could HTM help explain the high-order cognitive phenomena observed in humans (and orchestrated by the neocortex) in such a way that the Common Model of Cognition (CMC) can be enhanced, complemented, and extended? and vice versa, what kind of principles and main contributions could HTM borrow from CMC in order to allow building robust human-like artificial minds that are determined by behavioral, functional and neural constraints?

It is clear that HTM is neither a theory that explains how the entire brain works nor is interested in making machines that are human-like. Though, for those researchers who are interested in understanding how the decision-making process that continuously happens in our brains is not only the result of neocortical activity but also the result of the interaction of many non-cortical areas of the brain (like the basal ganglia, amygdala, hyppocampus, etc.), then it would be compelling to explore how HTM can be extended/integrated with other theories and computational frameworks.


You can’t run if you don’t know to walk. CMC is like seeing Usain Bolt on TV winning 100meters race and try to win the next Olimpic Gold medal not knowing how to walk.

Let us explain how the smallest shrew brain works (walk) and then we can scale up the problem (run). HTM is the way to start crawling … not the other way around.

1 Like

@Oscar_J_Romero - Welcome to the community!
I think that anyone here will agree that the scope of the HTM model is rather limited at this time.
I liken it to an intense focus on the transistor with the understanding that at some near future time this will be combined into useful larger structures like a computer. Numenta has stated that the focus on this integration into larger structures is anticipated in the future.

I know that you may be eager to see some sort of usable technology demonstration but I would like to point out that the PDP books came out in 1986 and the related technology models really did not hit their stride until the last decade or so. I don’t know if it will take 25 years but I think that it is unfair to expect instant results from the HTM model.

As far as the current progress of the deep learning community I think it is fair to point out that this extracts a single property of the brain (elaborations of layered perceptrons) and develops that mechanism to extract the useful property of data islands and manifold formation. The neuron used is a limited version of real neurons and fails to incorporate the learning mechanisms being explored in the HTM model.

Due to this simplification, the DL community has had to resort to heroic methods to load these structures with useful connection data. We know that the brain does most of this loading with far fewer presentations; HTM is an online system that learns with a comparable number of presentations and no requirement for supervision.

I anticipate a future fusion of the technologies to gain the advantages that each has to offer.

As far as a system levels approach to a high-level cognitive model needed to incorporate the sub-cortical structures - you are preaching to the choir here.

This thread incorporates most of my thinking along these lines, with various posts addressing different aspects of this very complex system.


It looks like Ben Goertzel’s team is already integrating HTM into a fledgling AGI architecture (OpenCog). They had to hack some things to get spatial and size generalization. I’m wondering if some of the HTM theorists know how the brain does this, because I’m sure it’s more elegant. See the video below. He starts talking about what they are doing with HTM at around 54:25.


A “more elegant” spatial and size generalization made me think of grid modules. Each 3 axis 2D representation almost doubles in resolution from the next.

By projection from one 2D size to the next grid modules may be representing objects and environments at essentially all possible visual sizes, not just the one being viewed at. In this case close up features would automatically be recalled, by what is seen from far away.

With grid modules not being accounted for in models yet, my best guess is that a more elegant way may be found there. It would also solve a signal ambiguity problem in my model that is the result of needing scales with less resolution, to best represent large areas to avoid.

I do not know of any evidence that would rule out this possibility, and felt it was worth mentioning, just in case this actually is what’s needed.

1 Like

This video is from 2013. I had not seen it, so thanks for posting! I looked into OpenCog awhile back and it seemed very disjointed. A kitchen sink of AGI ideas. Last I checked it was severely stalled, and they were focusing on blockchain instead.

Funny thing… the video above from Geotzel was posted about 6 months before NuPIC was open-sourced, so the version he was comparing to was a very old version.