Connectionists: AI at 50 videos...Explainable AI emerges naturally from Adaptive Resonance Theory

Sat Aug 29 05:08:17 EDT 2026

Dear Steve,

What DARPA seemed to want at that time for computer vision/object recognition is object prediction based on part verification – if a cat, have you verified its unique parts? That’s the kind of explanation they wanted. “Teaching” an object recognition system about these parts (to generalize, parts are just subconcepts) is not inconsistent with the way we teach humans. For example, a one- or two-year-old recognizes a car. But we teach them about parts much later - such as what a tire is, a door, steering wheel and so on. How a person’s internal biological model is adjusted from these “part teachings” I am not sure, but the identification and teaching (labeling) about parts is very much a part of human learning.

One might say that LLMs automate this type of learning, but we do provide labels in the descriptions of objects. So, that’s “teaching.”

When we started with multilayered models, the assumption was that at higher levels, there would be part abstractions such as those shown in the figure below - the dog face, legs, tail and so on. The problem was, we couldn’t find these abstractions at higher levels of the network. Even if you found them, we would need to assign labels or names to them – a dog face, legs, tail and so on. Once you get to assigning these labels, that’s “teaching.” There is a body of work in computer vision trying to build these abstractions inside CNN models. In our method, we also “teach” our deep learning models, but in a different way.

In general, “teaching” would be necessary whichever way you do it.

We call our approach “Explanation First, Model Next.” We have reversed the process of model building. Our testing shows gain in overall accuracy compared to some standard object detection models. This DARPA style approach also enables resistance against adversarial attacks. So substantial benefits with the Explanation First approach.

Best,
Asim

Asim Roy

Professor, Information Systems

Arizona State University

Asim Roy | iSearch (asu.edu)<https://isearch.asu.edu/profile/9973>

[cid:image003.png at 01DD3758.6623B9D0]

From: Grossberg, Stephen <steve at bu.edu>
Sent: Friday, August 28, 2026 8:01 AM
To: Asim Roy <ASIM.ROY at asu.edu>; Stephen José Hanson <jose at rubic.rutgers.edu>; Rothganger, Fred <frothga at sandia.gov>; Stevan Harnad <harnad at ecs.soton.ac.uk>
Cc: connectionists at mailman.srv.cs.cmu.edu; Grossberg, Stephen <steve at bu.edu>
Subject: Re: Connectionists: AI at 50 videos...Explainable AI emerges naturally from Adaptive Resonance Theory

Dear Asim,

Thanks for bringing up the DARPA Explainable AI program. I was invited to submit an article to a Special Issue about Explainable AI that was published when the DARPA program was announced. The editors of the Special Issue knew that Adaptive Resonance Theory, or ART, models automatically discover and incrementally learn over multiple learning trials to pay attention to the predictive, or explainable, critical feature patterns that control successful predictions and are used in learning by the adaptive weights in bottom-up adaptive filters and top-down expectations.

ART does this AUTOMATICALLY. You seem to have to use an external teacher to do it.

Here is the article:

Grossberg, S. (2020). A path towards Explainable AI and autonomous adaptive intelligence: Deep Learning, Adaptive Resonance, and models of perception, emotion, and action. Frontiers in Neurobotics, June 25, 2020.
https://www.frontiersin.org/journals/neurorobotics/articles/10.3389/fnbot.2020.00036/full

The article’s Abstract illustrates its explanatory range [boldface mine]:

  *   "Biological neural network models whereby brains make minds help to understand autonomous adaptive intelligence. This article summarizes why the dynamics and emergent properties of such models for perception, cognition, emotion, and action are explainable, and thus amenable to being confidently implemented in large-scale applications. Key to their explainability is how these models combine fast activations, or short-term memory (STM) traces, and learned weights, or long-term memory (LTM) traces. Visual and auditory perceptual models have explainable conscious STM representations of visual surfaces and auditory streams in surface-shroud resonances and stream-shroud resonances, respectively. Deep Learning is often used to classify data. However, Deep Learning can experience catastrophic forgetting: At any stage of learning, an unpredictable part of its memory can collapse. Even if it makes some accurate classifications, they are not explainable and thus cannot be used with confidence. Deep Learning shares these problems with the back propagation algorithm, whose computational problems due to non-local weight transport during mismatch learning were described in the 1980s. Deep Learning became popular after very fast computers and huge online databases became available that enabled new applications despite these problems. Adaptive Resonance Theory, or ART, algorithms overcome the computational problems of back propagation and Deep Learning. ART is a self-organizing production system that incrementally learns, using arbitrary combinations of unsupervised and supervised learning and only locally computable quantities, to rapidly classify large non-stationary databases without experiencing catastrophic forgetting. ART classifications and predictions are explainable using the attended critical feature patterns in STM on which they build. The LTM adaptive weights of the fuzzy ARTMAP algorithm induce fuzzy IF-THEN rules that explain what feature combinations predict successful outcomes. ART has been successfully used in multiple large-scale real-world applications, including remote sensing, medical database prediction, and social media data clustering. Also explainable are the MOTIVATOR model of reinforcement learning and cognitive-emotional interactions, and the VITE, DIRECT, DIVA, and SOVEREIGN models for reaching, speech production, spatial navigation, and autonomous adaptive intelligence. These biological models exemplify complementary computing, and use local laws for match learning and mismatch learning that avoid the problems of Deep Learning."

Chapter 5 in 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 a self-contained and non-technical overview and synthesis of my results about ART and Explainable AI.

The other chapters summarize neural network models of the main processes whereby our brains make our conscious minds in healthy individuals and clinical patients.

Best,

Steve

Stephen Grossberg
Wang Professor of Cognitive and Neural Systems
Director, Center for Adaptive Systems
Emeritus Professor of Mathematics & Statistics, Psychological & Brain Sciences, and Biomedical Engineering
Boston University
sites.bu.edu/steveg/
steve at bu.edu<mailto:steve at bu.edu>
http://en.wikipedia.org/wiki/Stephen_Grossberg
http://scholar.google.com/citations?user=3BIV70wAAAAJ&hl=en
https://sites.bu.edu/steveg/files/2021/08/Grossberg-CV-8-14-21.pdf
https://youtu.be/9n5AnvFur7I
https://www.youtube.com/watch?v=_hBye6JQCh4
https://www.amazon.com/Conscious-Mind-Resonant-Brain-Makes/dp/0190070552
https://www.amazon.com/Your-Creative-Brain-Consciously-Experience/dp/0198965370

From: Connectionists <connectionists-bounces at mailman.srv.cs.cmu.edu<mailto:connectionists-bounces at mailman.srv.cs.cmu.edu>> on behalf of Asim Roy <ASIM.ROY at asu.edu<mailto:ASIM.ROY at asu.edu>>
Date: Friday, August 28, 2026 at 3:51 AM
To: Stephen José Hanson <jose at rubic.rutgers.edu<mailto:jose at rubic.rutgers.edu>>; Rothganger, Fred <frothga at sandia.gov<mailto:frothga at sandia.gov>>; Stevan Harnad <harnad at ecs.soton.ac.uk<mailto:harnad at ecs.soton.ac.uk>>
Cc: connectionists at mailman.srv.cs.cmu.edu<mailto:connectionists at mailman.srv.cs.cmu.edu> <connectionists at mailman.srv.cs.cmu.edu<mailto:connectionists at mailman.srv.cs.cmu.edu>>
Subject: Re: Connectionists: AI at 50 videos...
Stephen,

You raise important issues. There are many ways of dealing with them. Here’s one of the ways. The image below shows what DARPA wanted for Explainable AI. Essentially, find the parts of an object for its prediction. So, for a cat, verify some of its unique parts like the fur, whiskers, and claws. Thus, in our explainable models, we “teach” an object detection model about these parts. What we get as outputs, for a cat, are predictions for the cat and its parts. There’s a layer of logic that uses these outputs and verifies the existence of the parts before “finally” predicting that there’s a cat in the image. In essence, it’s neuro-symbolic. One can do the symbolic part with a graphical model too. We can also predict that there’s a cat even if we don’t “see” all of the parts, such as when we just see its face and not the rest of the body. And, in a similar way, your “morning coffee” concept can be built from its parts that you mention.

There’s a long tradition in computer vision of finding parts of objects inside the model. Our initial conception was that the higher-level filters in a CNN correspond to certain abstractions such as a nose, eyes, ears for a human. That was the distributed representation but based on lower-level abstractions (parts of objects). But years of research showed that one could not find these kinds of lower-level abstractions in standard CNN models. So, what do you do? Well, one way is to force some of the filters to correspond to these parts. There’s a long history to this body of work in computer vision, including that of Hinton, to create abstractions within a model. Both ways, we are “teaching” the models about “low-level abstraction.” In a way, in both ways, it’s a hierarchical system. In both ways, we create a neuro-symbolic system.

By the way, abstractions are generalizations, they don’t code specific episodic memories. Your “table,” “coffee,” “cup” and all that are abstract notions.

Asim

Asim Roy

Professor, Information Systems

Arizona State University

Asim Roy | iSearch (asu.edu)<https://isearch.asu.edu/profile/9973>

[A screenshot of a computer  AI-generated content may be incorrect.]

From: Stephen José Hanson <jose at rubic.rutgers.edu<mailto:jose at rubic.rutgers.edu>>
Sent: Thursday, August 27, 2026 4:26 AM
To: Asim Roy <ASIM.ROY at asu.edu<mailto:ASIM.ROY at asu.edu>>; Rothganger, Fred <frothga at sandia.gov<mailto:frothga at sandia.gov>>; Stevan Harnad <harnad at ecs.soton.ac.uk<mailto:harnad at ecs.soton.ac.uk>>
Cc: connectionists at mailman.srv.cs.cmu.edu<mailto:connectionists at mailman.srv.cs.cmu.edu>
Subject: Re: Connectionists: AI at 50 videos...

Asim,

I sit here with my coffee neuron and my cup neuron creating a cup of coffee, and my table neuron holding the cup and coffee neurons representing a cup of coffee.  On the table, on the floor, noting of course the cup of coffee is not on the floor but on the table.   So this is a network of coffee, cups, tables, maybe coded by another neuron in a hierarchical network called morning coffee?    Well this is the typical problem with localist model of the brain.  How would it work exactly?   So far just representing a single episodic event requires hierarchical network of neurons that generalize to what exactly?    Do we need many such networks to represent all episodic memories of having a cup of coffee?

This is similar to the problem with the socalled "fusiform face area" which with some proper controls doesn't really exist (cf Hanson 2022).    What would a small blueberry are in fusiform gyrus be doing with all those faces we recognize?   A brain face cloud server?  Other accounts show that FG, LO, IN and PFC are at least involved in a network of areas (not neurons) in recognizing a face.   But computationally what's the actual underlying functions?

The shift to "neurons represent symbols' is not helpful to explain meaning.

 Remember, we have billions of neurons and 100 trillions connections between them.

I'm not saying symbols aren't important and Harnad's Grounding arguments are becoming more critical (and need to be made more explicit)  as LLMs start to be embedded in robots.     Therefore an arguably simpler task, is to identify "symbols" in LLMs since we have no idea how they work either. Are symbols  in their "neurons"?

Cheers

Stephen
On 8/24/26 22:56, Asim Roy wrote:

  1.  Plate (2002): “Another equivalent property is that in a distributed representation one cannot interpret the meaning of activity on a single neuron in isolation: the meaning of activity on any particular neuron is dependent on the activity in other neurons (Thorpe, 1995).”
  2.  Thorpe (1995, p. 550): “With a local representation, activity in individual units can be interpreted directly … with distributed coding individual units cannot be interpreted without knowing the state of other units in the network.”
  3.  Elman (1995, p. 210): “These representations are distributed, which typically has the consequence that interpretable information cannot be obtained by examining activity of single hidden units.”

1. Elman J. (1995). Language as a dynamical system, in Mind as Motion: Explorations in the Dynamics of Cognition, eds Port R., van Gelder T. (Cambridge, MA: MIT Press), 195–223.
2. PlateT. (2002). Distributed representations, in: Encyclopedia of Cognitive Science, ed Nadel L. (London: Macmillan), 2.
3. ThorpeS. (1995). Localized versus distributed representations, in The Handbook of Brain Theory and Neural Networks, ed Arbib M. (Cambridge, MA: MIT Press), 550.

In that discussion with Horace Barlow and others, interpretation and meaning of the activations that responded only to certain specific stimuli was the main source of discomfort to connectionists including Walter Freeman. Hence Walter asking Rodrigo directly whether those activations had meaning and interpretation.

Asim

Asim Roy

Professor, Information Systems

Arizona State University

Lifeboat Foundation Bios: Professor Asim Roy<https://urldefense.com/v3/__https:/lifeboat.com/ex/bios.asim.roy__;!!IKRxdwAv5BmarQ!NHg1P4byXZZmMFLMgr8A4dQNxSlqyvsqWef4wrPXtCXR0f0bo0eMm-rIF8G6ZV0$>

Asim Roy | iSearch (asu.edu)<https://isearch.asu.edu/profile/9973>

From: Rothganger, Fred <frothga at sandia.gov><mailto:frothga at sandia.gov>
Sent: Monday, August 24, 2026 8:46 AM
To: Asim Roy <ASIM.ROY at asu.edu><mailto:ASIM.ROY at asu.edu>
Cc: connectionists at mailman.srv.cs.cmu.edu<mailto:connectionists at mailman.srv.cs.cmu.edu>
Subject: AI at 50 videos...

Asim,

It makes sense for you to put "meaning" in scare-quotes, since it is another ill-defined term. IMHO, the important thing is adaptive behavior. That a particular neuron fires when the organism is exposed to a stimulus is significant because that spike does work on other parts of the system, ultimately shaping behavior. The notion of "meaning" may just be a convenience for us as observers discussing the system.

-- Fred

________________________________
From: Connectionists <connectionists-bounces at mailman.srv.cs.cmu.edu<mailto:connectionists-bounces at mailman.srv.cs.cmu.edu>> on behalf of Asim Roy <ASIM.ROY at asu.edu<mailto:ASIM.ROY at asu.edu>>
Sent: Friday, August 21, 2026 10:47 PM
To: KENTRIDGE, ROBERT W. <robert.kentridge at durham.ac.uk<mailto:robert.kentridge at durham.ac.uk>>; Gary Marcus <gary.marcus at nyu.edu<mailto:gary.marcus at nyu.edu>>; Stephen José Hanson <jose at rubic.rutgers.edu<mailto:jose at rubic.rutgers.edu>>; Hava Siegelmann (hava.siegelmann at gmail.com<mailto:hava.siegelmann at gmail.com>) <hava.siegelmann at gmail.com<mailto:hava.siegelmann at gmail.com>>; Juergen Schmidhuber <juergen.schmidhuber at kaust.edu.sa<mailto:juergen.schmidhuber at kaust.edu.sa>>; Ali Minai <minaiaa at gmail.com<mailto:minaiaa at gmail.com>>
Cc: director at inc.ucsd.edu<mailto:director at inc.ucsd.edu> <director at inc.ucsd.edu<mailto:director at inc.ucsd.edu>>; connectionists at mailman.srv.cs.cmu.edu<mailto:connectionists at mailman.srv.cs.cmu.edu> <connectionists at mailman.srv.cs.cmu.edu<mailto:connectionists at mailman.srv.cs.cmu.edu>>
Subject: [EXTERNAL] Re: Connectionists: FW: AI at 50 videos...

By the way, most of these microelectrode-based studies, starting with Horace Barlow, are single cell studies. And the single cell studies are the ones that have won Noble prizes because of the insights they produced. The overall evidence from these studies is not just about grandmother cells, but about single cells encoding abstractions and “meaning.” Reddy and Thorpe (2014) conclude:

“In conclusion, evidence is accumulating that “concept cells” carry high-level, abstract stimulus information.”

Single cells having “meaning” was (and still is) at the heart of discussion at that time because it contradicts the fundamental population coding idea where single neurons by themselves have no meaning, they only have meaning collectively. And Walter Freeman at that time could not accept the idea that single cells have meaning. We were supposed to have a public debate about this at IJCNN 2011 in San Jose. Walter posed the question directly to Rodrigo Quian Quiroga, who conducted many of these experiments under the supervision of Itzhak Fried and Christof Koch. Walter also knew Rodrigo, having been co-authors or co-editors of a book. Rodrigo came through at the last minute before the public debate, acknowledging that the single cells in his studies indeed had “meaning.”

“Meaning” and abstraction is at the heart of these debates, not grandmother cells. And finding “meaning” in activations of single cells directly contradicts the population coding hypothesis and supports instead the symbol system hypothesis.

Asim Roy

Professor, Information Systems

Arizona State University

Lifeboat Foundation Bios: Professor Asim Roy<https://urldefense.com/v3/__https:/lifeboat.com/ex/bios.asim.roy__;!!IKRxdwAv5BmarQ!NHg1P4byXZZmMFLMgr8A4dQNxSlqyvsqWef4wrPXtCXR0f0bo0eMm-rIF8G6ZV0$>

Asim Roy | iSearch (asu.edu)<https://isearch.asu.edu/profile/9973>

--
[cid:image002.jpg at 01DD3755.3D4B3630]
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.srv.cs.cmu.edu/pipermail/connectionists/attachments/20260829/ef7f8265/attachment-0001.html>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: image001.jpg
Type: image/jpeg
Size: 45946 bytes
Desc: image001.jpg
URL: <http://mailman.srv.cs.cmu.edu/pipermail/connectionists/attachments/20260829/ef7f8265/attachment-0002.jpg>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: image002.jpg
Type: image/jpeg
Size: 57383 bytes
Desc: image002.jpg
URL: <http://mailman.srv.cs.cmu.edu/pipermail/connectionists/attachments/20260829/ef7f8265/attachment-0003.jpg>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: image003.png
Type: image/png
Size: 333780 bytes
Desc: image003.png
URL: <http://mailman.srv.cs.cmu.edu/pipermail/connectionists/attachments/20260829/ef7f8265/attachment-0001.png>