Connectionists: AI at 50 videos...: Almost 600 neural network modeling articles since 1957 model the main processes whereby our brains make our minds

Thu Aug 27 19:46:30 EDT 2026

Dear Stephen. ..  Thank you for taking the time to go through all of this and provide such detailed responses with references.  It sounds like you agree that we learn both concepts and symbols that refer to concepts.   And you are most certainly correct about building on the findings of others versus repeating such work.
.. Danny

Sent from my Bell Samsung device over Canada’s largest network.
________________________________
From: Grossberg, Stephen <steve at bu.edu>
Sent: Thursday, 27 August 2026 12:00:19
To: Danny Silver <danny.silver at acadiau.ca>; Barak A. Pearlmutter <barak at pearlmutter.net>
Cc: connectionists at mailman.srv.cs.cmu.edu <connectionists at mailman.srv.cs.cmu.edu>; Grossberg, Stephen <steve at bu.edu>
Subject: Re: Connectionists: AI at 50 videos...: Almost 600 neural network modeling articles since 1957 model the main processes whereby our brains make our minds

CAUTION: This email comes from outside Acadia. Verify the sender and use caution with any requests, links or attachments.

Dear Danny,

It’s good to hear from you. You raise several issues. I reply to them in turn.

In fact, ALL of the issues that you have described have been clarified over the years by biological neural networks that I and my colleagues have developed and used to provide principled explanations and quantitative computer simulations of large amounts of psychological and neurobiological data.

You write below:

"Young children develop conceptual representations of things like mother, food they like, asking for more, or things that scare them - long before they have symbols for these concepts.  But once they have learned such symbols (hand gestures or words) they can use them to communicate about these concepts to others, in albeit limited ways."
+++++++++++++++
You might find the following two articles relevant here:

Grossberg, S. (2023). How children learn to understand language meanings: A neural model of adult–child multimodal interactions in real-time. Frontiers in Psychology, August 2, 2023. Section on Cognitive Science, Volume 14.
https://www.frontiersin.org/journals/psychology/articles/10.3389/fpsyg.2023.1216479/full

Grossberg, S. (2025). Neural network models of autonomous adaptive intelligence and artificial general intelligence: How our brains learn large language models and their meanings. Frontiers in Systems Neuroscience, July 29, 2025, Volume 19.
https://www.frontiersin.org/journals/systems-neuroscience/articles/10.3389/fnsys.2025.1630151/full

These articles clarify how children incrementally LEARN symbols and language utterances that have perceptual and affective MEANINGS by interacting with caregivers in real time.

I also provide a self-contained and non-technical explanation of how this works in my 2026 book:

YOUR CREATIVE BRAIN AND AI:
How We Learn and Consciously Experience
ART, MUSIC, and MEANING
https://www.amazon.com/dp/0198965370?lv=shuf&channelId=500&plpRedirect=mhFallback

+++++++++++++++++
You also wrote below:

"Many lower-order animals seem to function in the world very well - learning about and using concepts - without ever having need for symbols.  So symbols are not required to think, but they make explaining our thinking to others possible (kind of).”

Primitive organisms like crustacea can function using stereotyped circuits, including what I have called context-sensitive avalanches; see Figure 4 in:

Grossberg, S. (1970). Some networks that can learn, remember, and reproduce any number of complicated space-time patterns, II. Studies in Applied Mathematics, 49, 135-166.
https://sites.bu.edu/steveg/files/2016/06/Gro1970SiAM.pdf

or Chapter 5 in:

Grossberg, S. (1974). Classical and instrumental learning by neural networks. In R. Rosen and F. Snell (Eds.), Progress in theoretical biology. New York: Academic Press, pp. 51-141.
https://sites.bu.edu/steveg/files/2016/06/Gro1974ProgressTheorBiol.pdf

Chapter 13 of my 2021 Magnum Opus

CONSCIOUS MIND, RESONANT BRAIN: HOW EACH BRAIN MAKES A MIND
https://www.amazon.com/dp/0190070552?lv=shuf&channelId=500&plpRedirect=mhFallback

provides an evolutionary account of how avalanches have developed through phylogeny to replace stereotyped performance with learned sequential performance that is sensitive to multiple types of environmental feedback, leading to the cognitive and emotional circuits in human brains.

These circuits do not, however, enable a primitive organism to “think”. That requires brain circuits that support attention, learning, cognition, and action, which my other article model, including articles about Adaptive Resonance Theory.

Another way that animals “function in the world very well” is illustrated by terrestrial animals who can learn to navigate by using spatial maps and cognitive-emotional interactions. Many of these animals also use learned recognition categories to decide in what direction to navigate to achieve effective foraging.

For example,

Browning, A., Grossberg, S., and Mingolla, M. (2009). Cortical dynamics of navigation and steering in natural scenes: Motion-based object segmentation, heading, and obstacle avoidance. Neural Networks, 22, 1383-1398.
https://sites.bu.edu/steveg/files/2016/06/BroGroMinNN2009.pdf

Yet another article shows how prefrontal cortex can interact with multiple brain regions to enable humans and higher-animals to plan and act to realize valued goals at socially-appropriate times.

For example,

Grossberg, S. (2018). Desirability, availability, credit assignment, category learning, and attention: Cognitive-emotional and working memory dynamics of orbitofrontal, ventrolateral, and dorsolateral prefrontal cortices. Brain and Neuroscience Advances, May 8, 2018.
https://journals.sagepub.com/doi/full/10.1177/2398212818772179
+++++++++++
You write below:

"Stephen, I can see how an ART network can develop conceptual feature representations, say for the concept <cat> from photos, before the network becomes aware of a symbol for <cat>, such as the spoken word “cat".   The associated recognition category may simply not have a label.   But how is this conceptual representation later bound to a symbolic representation for the evert symbol “cat”.  This is particularly messy when one considers there are all kinds of cats - ones you can pet, one that can eat you.

"Perhaps through the simultaneity of both seeing a cat and also hear the word “cat”.  Hmmm … Very interesting.”

Danny, the above comments require several different interacting brain mechanisms to be explain and modeled.

You write:

"how an ART network can develop conceptual feature representations, say for the concept <cat> from photos, before the network becomes aware of a symbol for <cat>, such as the spoken word “cat".   The associated recognition category may simply not have a label.”

The ARTMAP family of neural networks model how this works. They can incrementally learn in response to arbitrary sequences of unsupervised and supervised learning trials. For example:

Carpenter, G.A., Grossberg, S., Markuzon, N., Reynolds, J.H., and Rosen, D.B. (1992). Fuzzy ARTMAP: A neural network architecture for incremental supervised learning of analog multidimensional maps. IEEE Transactions on Neural Networks, 3, 698-713.
https://sites.bu.edu/steveg/files/2016/06/CarGroMarRey1992IEEETransNN.pdf

You also write: “there are all kinds of cats - ones you can pet, one that can eat you”. The process of vigilance control modulates ART category learning, with low vigilance leading to the learning of general, or abstract, recognition categories and high vigilance learning to the learning of specific, or concrete recognition categories, including exemplars.

An automatic process called match tracking leads to minimax learning that conjointly maximizes category generality while minimizing predictive error.

If vigilance is too low to predict the right answer when presented with one kind of cat, match tracking is triggered, thereby leading to hypothesis testing or memory search, for the other kind of cat. If the other kind of cat has not yet been learned, search ends an a novel category is learned to recognize the other kind of cat.

Here is the first article that proved theorems for how vigilance control does this:

Carpenter, G.A., and Grossberg, S. (1987). A massively parallel architecture for a self-organizing neural pattern recognition machine. Computer Vision, Graphics, and Image Processing, 37, 54-115.
https://sites.bu.edu/steveg/files/2016/06/CarGro1987CVGIP.pdf

This 1987 article predicted the existence of vigilance control. The following article shows that it exists, and explains psychological, anatomical, neurophysiological, biophysical, and biochemical data about how it works:

Grossberg, S. and Versace, M. (2008). Spikes, synchrony, and attentive learning by laminar thalamocortical circuits. Brain Research, 1218, 278-312.
https://sites.bu.edu/steveg/files/2016/06/GroVer2008BR.pdf
See Figure 5.
+++++++++++++++++++++

More generally, I have published almost 600 archival neural network modeling articles since 1957 that are downloadable from sites.bu.edu/steveg. They develop biological neural network models of the main processes whereby our brains make our conscious minds in health individuals and clinical patients.

As I note on my web page, with over 100 gifted PhD students, postdocs, and faculty, I have discovered and led the development of  "brain models of vision and visual object recognition; audition, speech, and language; development; attentive learning and memory; cognitive information processing and social cognition; reinforcement learning and motivation; cognitive-emotional interactions; navigation; sensory-motor control and robotics; and mental disorders. These models involve many parts of the brain, ranging from perception to action, and multiple levels of brain organization, ranging from individual spikes and their synchronization to cognition. Many of these projects are done in collaborations with PhD students, postdoctoral fellows, and faculty. I also collaborate with experimentalist colleagues to design experiments to test theoretical predictions and fill conceptually important gaps in the experimental literature, carry out analyses of the mathematical dynamics of neural systems, and transfer biological neural models to applications in neuromorphic engineering and technology.”

I recommend to everyone that, before trying to develop a neural model of some brain process, you read the article titles and abstracts on sites.bu.edu/steveg. It would be a waste of your time and talent to just reinvent the wheel.

It would be much more productive to build on the secure foundation of what is already known.

Best,

Steve

Get Outlook for Mac<https://aka.ms/GetOutlookForMac>
From: Danny Silver <danny.silver at acadiau.ca>
Date: Wednesday, August 26, 2026 at 9:54 PM
To: Grossberg, Stephen <steve at bu.edu>; Barak A. Pearlmutter <barak at pearlmutter.net>
Cc: connectionists at mailman.srv.cs.cmu.edu <connectionists at mailman.srv.cs.cmu.edu>
Subject: Re: Connectionists: AI at 50 videos...

Dear Barak and Hello Stephen (great to speak with you again).
I have attached my original email below for Stephen, as the last word from the connectionist mail list was that my original email  was pending approval by the moderator.  So not sure how Stephen saw your response but perhaps not my original email.
Anyway .. the miracles of the web.

Barak .. I like your paraphrase - you definitely get the major idea of the paper.  Young children develop conceptual representations of things like mother, food they like, asking for more, or things that scare them - long before they have symbols for these concepts.  But once they have learned such symbols (hand gestures or words) they can use them to communicate about these concepts to others, in albeit limited ways.   Many lower-order animals seem to function in the world very well - learning about and using concepts - without ever having need for symbols.  So symbols are not required to think, but they make explaining our thinking to others possible (kind of).

But please note,  there is a more subtle second idea that you may have missed.  Animals (particularly humans) who use symbols to communicate externally with each other also gained the advantage of creating an additional constraint (a bias) for learning and reasoning about the world.  As I say in the original email  -  what we learn and reason about is to some extent constrained/shaped by the “lexicon” of symbols (and their related concepts) that we have already created or learned from others.   Humans initially developed the ability to manipulate symbolic representations to communicate their thoughts to each other - it was necessary to survive.   But then a beautiful thing happened,  we started using the symbols and this language of thought to constrain/rationalize our thinking - or at least, at the best of times we do so {:]).

Stephen, I can see how an ART network can develop conceptual feature representations, say for the concept <cat> from photos, before the network becomes aware of a symbol for <cat>, such as the spoken word “cat".   The associated recognition category may simply not have a label.   But how is this conceptual representation later bound to a symbolic representation for the evert symbol “cat”.  This is particularly messy when one considers there are all kinds of cats - ones you can pet, one that can eat you.
Perhaps through the simultaneity of both seeing a cat and also hear the word “cat”.  Hmmm … Very interesting.

… Danny

From: Grossberg, Stephen <steve at bu.edu>
Date: Wednesday, August 26, 2026 at 3:58 PM
To: Barak A. Pearlmutter <barak at pearlmutter.net>, Danny Silver <danny.silver at acadiau.ca>
Cc: connectionists at mailman.srv.cs.cmu.edu <connectionists at mailman.srv.cs.cmu.edu>
Subject: Re: Connectionists: AI at 50 videos...

CAUTION: This email comes from outside Acadia. Verify the sender and use caution with any requests, links or attachments.
Evert recognition category provides a “ symbolic “ representation of the distributed feature pattern that it represents. A feature-category resonance links category and features into a bound state that enables conscious recognition of the features. My 2021 OUP book explains this in detail and provides lots of experimental support for it.

Get Outlook for iOS<https://aka.ms/o0ukef>
________________________________
From: Connectionists <connectionists-bounces at mailman.srv.cs.cmu.edu> on behalf of Barak A. Pearlmutter <barak at pearlmutter.net>
Sent: Wednesday, 26 August 2026 13:04:32
To: Danny Silver <danny.silver at acadiau.ca>
Cc: connectionists at mailman.srv.cs.cmu.edu <connectionists at mailman.srv.cs.cmu.edu>
Subject: Re: Connectionists: AI at 50 videos...

Dear Danny,

If I might paraphrase that work, it seems to me the idea is that
"subsymbolic" signals are used for local processing, while symbolic
representations are used for longer distance communication, with
things becoming more symbolic the longer and narrower the channel:
from one brain region to another, from the brain to its future self,
from one brain to another.

That certainly seems consistent with the information bottleneck of
Tishby et al, where distributions fragment into clustered
representations as the bottleneck becomes more severe.

My only issue is that, if you measure things in a conventional
computer doing symbol processing, the individual bits on wires can be
seen as subsymbolic. What is bit 7 of register 4? What is bit 11 of
the address bus? It is not until you agglomerate things just right, at
just the right level of abstraction, that the crisp symbolic nature
of, say, a compiler running on a simple RISC, becomes apparent. This
subproblem seems to subsume the whole question, because it's basically
"figure out how it all works and then we can talk about whether there
are symbolic representations." This might not obviate the distinction
between symbolic and subsymbolic, but I'd argue that it does make it
uninteresting.

Cheers,

--Barak.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.srv.cs.cmu.edu/pipermail/connectionists/attachments/20260827/86adb8e2/attachment.html>