The first reading in Andy Lattal’s verbal behavior course is:

Here is where I take leave of the usual psychologist’s comparison of viewpoints (or take leave of my senses: you be the judge). I don’t really care much what Chomsky’s viewpoint is, per se.

What I will comment on is an issue raised on page 40: “Just as cells in the visual cortex are organized in special ways not characteristic of cells controlling, say, digestion, …”. It is followed by this quote: “The language system, the visual system, and no doubt other systems, he asserts, may be modular.”

For a software developer, modularity is a very cool thing. In this case, it would provide fracture lines along which to break down the problem of using, say, a neural network, to model behavior. Any major functionality that we can cleanly factor out into a separate object(s) will make design and implementation much simpler. Same as for any complex task: break it down into a set of simpler tasks, then break them down … ad nauseum.

Is there utility in Chomsky’s notions of modularity?

Visual cortex

The visual cortex quote implies that modularity of the visual systems is settled science. And it does indeed seem to be. If you Google something along the lines of 'specialized cells and functions of visual cortex‘ you will get an enormous number of hits that more or less agree in their basics. There is also a book to which Dave Palmer contributed a chapter (Dave’s chapter is: Chapter 14: Selectionist constraints on neural networks, pp 263-282):

The relevant chapter regarding the visual cortex is:

  • Chapter 9: Inferotemporal Cortex and Object Recognition, pp 160-188.

It does seem that there is specialization. The specialized region does “pre-processing” that recognizes edges, orientations, movements, etc. But the units recognized are rather limited. The region outputs to the sensory area, in general, where most of the object recognition occurs. And it is subject to learning.

Do not remember where I read it, maybe in the above chapter, but it appears that this “pre-processor” gets “programmed” very soon after birth. The kind of pre-processing, the kinds of detections, that are performed by the visual cortex are at least partially determined by the visual environment. It is not readily changeable from that point forward, but if the organism moves to a very different visual environment with accompanying changes in contingencies, the visual cortex will gradually adapt to provide nearly instantaneous detection of critical visual features. Could be a life and death matter, ya think?

Does that mean that in a behavior modeling system, one should implement a specialized “visual cortex” emulator? Maybe. However, there are shortcuts. As always, if there is existing software that will perform the task adequately, one should evaluate existing system. There are many.

For example, your Apple devices contain vision processing WAY beyond what is required for emulating a visual cortex. I would investigate this functionality first, and evaluate which functionality is useful, and which is going too far beyond visual cortex functionality. Then write a “wrapper” around the appropriate functionality to create a visual cortex (and only a visual cortex). This code will be much faster than anything I could write, and do a better job:

Language cortex?

I think not. At least nothing particularly useful to me as a behavior modeler.

My searches regarding specialized brain areas for handling verbal behavior failed to retrieve findings similar to those found for the visual cortex. It would appear that handling of verbal behavior does not require such specialized pre-processing, and is handled pretty much as “just another behavior”. What little specialization there might be is at the level of left hemisphere versus right hemisphere. There is still argument about the level of any specialization, but the lack of agreement and clarity given all of the modern tools for investigating the activation of brain areas during particular activities argues that either (a) we are still not there in terms of tools, or (b) there is nothing to find.

In any case, there don’t appear to be any “fracture lines” here that would be useful in factoring out functionality to place into a specialized code object for handling verbal behavior. The processing of language appears to be so diffuse as for it to appear no different than any other behavior. So sad, so sorry.

This also means that unlike the case of the visual cortex, there really is no extant, freely available, “language cortex” software. What is available is software such as I mentioned before for traditional NLP, as well as higher powered “pattern recognition” language software such as Google’s BERT. None of it is really what we want if we are talking about sensitivity to factors we commonly associate with behavior analysis. Useful, yes, but not so much for modeling behavior from a behavior analytic perspective.

But if you are interested in traditional NLP (and it is very useful for many tasks), here is Apple’s take on it. You will see a set of NLP functionality nearly the same as in previous links:

Nature is the model

On the other hand, there is a useful specialization here. If verbal behavior is just behavior, then we can focus on the general learning functionality that also applies to all other behaviors.

Unlike computer science, nature does not generally seem to overspecialize. Or at least such overspecialization generally does not survive, thus we see little of it. On the other hand, “computer sciency” neural nets, and machine learning in general, are prone to “overfitting”. They do not generalize. Lack of generality seems to be less of a problem for bio-plausible neural net models, but bio-plausible models are not as fast and efficient for the kinds of tasks for which neural nets are normally used these days, thus they are not used (or not so much).

The overfitting issue arises periodically in the AI world. AI reaches dead ends with regard to what can be achieved with a particular methodology. We saw this in a previous cycle with “expert systems”. Expert systems are useful. They still exist as decision support software. But they are not nearly as general purpose as was once hoped. We are reaching that point in the current cycle.

When AI workers begin recognizing that the end is nigh, they generally go back to nature. They find some process that has evolved over billions of years, extract some useful algorithms, and off we go into the next cycle. Sounds silly. It is not. There are technical limitations. The limitations can take decades to overcome.

For example, we routinely use the current crop of AI methods on a daily basis: they are baked into the little mobile devices in our pockets. Only a few years ago such functionality would have required a room-sized super computer. Yep, you carry around a 1990s super computer in your pocket. You are likely carrying around more compute capacity in your pocket than existed in the entire world when I was born (and yes, there were computers then).

AI workers are beginning to dip into natural phenomena for clues:

Interesting notion: “noise” is very important. Ask yourself, where does new behavior come from? Could there be any new behavior for the environment to select without variability?

Anyway, I think we are approaching a period in which “bio-plausible” neural nets will become important. Which means that there is an opportunity for folks who have a deep understanding of the behavior of natural biological systems that include organic neural nets being modeled by the bio-plausible nets. You know, folks such as behavior analysts.

It is becoming a hot topic in the AI world. One sample article:

The article mentions “spiking models“. Another hot topic. See:

In reading through the book (no, I did NOT work through the math), I came to the conclusion that the Donahoe, Burgos, Palmer (DBP) model closely corresponds to the spike response model, in particular the SRM0 variant discussed in the book. José Burgos strongly disagrees, but skimming through this book might help one to understand the workings of the DBP model, anyway.

Your day is coming. Soon.

My intuition (meaning the above facts don’t fully support my conclusion) is that the day is coming very soon when behavior analysts COULD be in demand for understanding the behavior of neural nets that more closely model nature. If the technology progresses in the direction it seems to be going.

But more importantly, we see where the parade is going. We need to position ourselves in front of it and become a leading force. At some point, artificial organisms containing more naturalistic neural nets become just another organism. Who understands the behavior of organisms better than behavior analysts? Let me think. Oh, nobody.

So: how do we treat verbal behavior differently from other fields? What parts of it are amenable to encapsulation in software? What might computational behavior analysis be?