I am curious, as a thought experiment, or something more… What would it mean to introspect a cortical column? Is that something we do regularly in our lives? Is that part of our subjective human experience? What can we learn from that?
I am new to Thousand Brains. I suppose that a cortical column may be assigned to a soda can, or some other object in the world around us, including all of the angles by which we might perceive it, and the qualities it may have.
But I myself think that a cortical column may just as well be assigned to an abstract concept. And indeed, that abstract concept may be more revealing. The ways that we can conceive that concept may reveal the “user requirements” that explain why there should be six cortical layers.
I have been developing the idea that subjectively we experience our lives as the interplay of three levels of awareness:
- an answering mind that unconsciously knows
- a questioning mind that consciously does not know
- an investigating mind that aligns the other two minds, so that we have the same information in two different forms - what we know and what we do not know, and then selects which mind to implement
At Theory Translator, I and my friends have collected more than 300 examples from the history of world culture. I also describe the three minds in my paper An Allegory: The Solipsistic Self as the Hamiltonian of a Noninteracting Fermion.
The answering mind, the unconscious, can be thought of as a backwards looking Passive Inference, summing the knowledge of 100 billion neurons / answers, enmeshed with the world. The questioning mind, the conscious, can be thought of as a forwards looking Active Inference, divorced from the world, expressed by a generative model with some 100 thousand concepts / words / variables / slots / questions / black boxes (or cortical columns).
In this view, each cortical column expresses a concept. I wrote a proposal for an Active Inference Institute Research Fellowship entitled Modeling Subjective Human Experience as an Interplay of Passive, Active and Willful Inference. I propose to study the Encyclopedia of Triangle Centers as a corpus to sort out triangle geometry in terms of actions (by which issues matter), conceptions (triangle centers - by which meaning arises) and equivalences (by which events happen). The language by which we interact with an arbitrary triangle may well be the language by which we interact with anything. And that language may be encoded in a cortical column.
Interesting stuff, for sure. FWIW, my impression is that each cortical column is thought to develop a lot of different models, rather than just one. So, for example, the column handling input for a particular patch of the retina might have models for cups, handles, and logos, but also for bicycles, cars, wheels, etc. I don’t think this negates any of your arguments, however.
2 Likes
Rich, thank you! I suppose a lot of the research comes from study of vision and the retina, and informs our understanding accordingly.
I alao find your previous work and ideas very interesting. Regarding my understanding of cortical columns (CC), I think that it is important to note that they exist within a cognitive hierarchy amongst themselves. So which particular object or concept a cortical column is internally representing and modeling depends on its level in the hierarchy. And in my opinion it is the TBT voting process that determines the winner CCs. So there is a lot going on both within a CC as well as outside the CC in terms of selective competition. And all this within a hierarchical framework of CCs. And this framework does not organize itself into a single semantic hierarchy, but instead into multiple hierarchies, as many as experienced events require to model reality.
Joseph-Anthony, thank you for your reply. I’m interested to learn more about the cognitive hierarchy. Who could recommending some reading materials? I have this hunch that the six cortical layers could be expressing six kinds of visualizations. I once visited a hermit, Kęstas Augutis, who had a vision that the only requirement to graduate high school should be to write three books, demonstrating a mastery of three kinds of thinking: A diary (or blog) in sequential form, a thesaurus organized as a tree of concepts, and an encyclopedia written as a web of articles. I wondered if those were the only ways of structuring thoughts and so I collected about fifty examples. Curiously, sequences and hierarchies and networks are never used in isolation but always in pairs, with one restructuring the other, yielding six kinds of visualizations. At the World Congress of Universal Logic, I gave a presentation: Visualization as Restructuring and thus a Source of Logical Paradox. Since then I have wondered if the six cortical layers may correspond to these six kinds of visualization. I wonder where I could learn more about what is known about the six layers.
1 Like
Very fascinating and thought provoking ideas. I will need more time to read and reflect on these concepts. Thanks for sharing!
I personally cannot see any connection to the six layers (which are actually more than six according to newer research). But I certainly cannot discard a connection, because a lot is still to be learned about their full functionality. Known functions of some some layers are things like inhibition of neighboring columns to achieve sparcity in the activation patterns. Layer 4 is known to receive direct sensory input from sensory organs. In the HTM Forum you can read a lot more about this. It is also sponsored by Numenta.
SDRs = Sparse Distributed Representations. These are at the core of the HTM computational theory of the neocortex. You need to read about SDR first and then the layered wiring in the cortical columns will make more sense. But this does not discard that they fulfill further attentional functions. But I personally think that those mechanisms you are searching for are not implemented in the 6+ layers within the columns but rather in the hierarchical wiring interconnecting regions of columns. And that is in the realm of this TBT community.
1 Like
It’s conceivable that a cortical column could introspect (i.e., model) its own activities, but that seems to me like a fairly specialized task. That said, I’m quite interested in the provision of facilities within Monty that would allow collection (and ultimately, understanding) of its activities.
So, for example, it should be possible (nay, trivial) to get information on the state of a module (e.g., LM, SM). One way to do this would be to build in something like GraphQL:
GraphQL is a data query and manipulation language that allows specifying what data is to be retrieved (“declarative data fetching”) or modified. A GraphQL server can process a client query using data from separate sources and present the results in a unified graph.
This would let any interested party (e.g., a developer) ask about any aspect of the module’s current model(s). And, by extension, develop an understanding of the activities of suites of modules.
I’d also like models and modules to be “tagged” with information that indicates their nature. So, for example, it should be possible to find out the set of SMs that are feeding into a given LM. Alternatively, if a suite of modules has been trained on certain real-world objects, it would be useful to have that information available (e.g., to build libraries of models).
To be clear, I’m pretty sure that these sorts of facilities aren’t present in the biological cortex. However, as a software developer, I’m in favor of anything that makes the systems’ operations more transparent, understandable, etc.
1 Like
I certainly support your views on the benefits (if not strickt necessity) for good computational analytical transparency. I also trust that this TBP community also generally agrees and has at least some tools and techniques in place. Your technical proposal seems to be very comprehensive and probably extends beyond what is use today. (I am not best informed on this matter). Graph DB technology and proper logging techniques can go a long way in helping to understand model behavior and performance.
1 Like