Research Compatible with PCT versus Predictive Processing

This is to point out that the research summarized by Huijeong Jeong et al, “Mesolimbic dopamine release conveys causal association”, Science, 378, 6626, 23 December 2022, 1294 (whose full article is available via the link below) seems to support Bill Power’s PCT explanation of what is happening neurally as opposed to that of Predictive Processing.******

“These models make different predictions about mesolimbic dopamine, a critical controller of learning.” A very nice piece of work!

The mesolimbic pathway transports dopamine from the ventral tegmental area (VTA) in the midbrain to several parts of the brain, relevantly in this case to the nucleus accumbens and amygdala.

Could you elaborate on this a bit. I’m not sure I understand what the researchers (or the subjects) did or why you think what the researchers found supports “Powers’ explanation of what is happening neurally”. What is Powers’ explanation of what is happening neurally in this situation. What is the situation of the subjects? I presume it’s a conditioning task. Is that right?

Best, Rick

As I read it, the research supports the view that organisms construct input functions to perceive causal relationships, and the alternative view is based on a theory of processing prediction error.

The article’s authors 11 experiments did not support hypotheses consistent with Predictive Processing whereas they did support hypotheses based on their concept/model (called ANCCR for adjusted net contingency for causal relations). In turn, their concept and model seem consistent with Bill Powers PCT-based thinking of memory, learning, and stored perceptions of past events – i.e., as related to the article’s meaningful cues (that they call stimuli) and related learning associated with delayed meaningful outcomes (what the authors call “rewards” but PCT would consider to be indicators of “error reduction”).

PS: In case you haven’t seen it Rick, I’ve sent you a copy by email of the full Research Article so you can have a look and reach your own conclusions about possible compatibility with PCT and Bill’s thinking about neural functioning.

I did get the Jeong et al article. I still don’t see why you think this research is compatible with PCT versus Predictive Processing. Indeed, I don’t see how one can tell whether it’s compatible with PCT or not. For one thing, it’s not clear to me how their dependent variable – dopamine release – fits into PCT. Apparently the two models described in the Joeng et al paper make different predictions about how much dopamine will be released during the course of a discriminative conditioning task. So dopamine must have something to do with learning in which case the aspect of PCT that applies to this research is the reorganization model. And it would be the aspect of the reorganization model that explains why the rats in this experiment learn that a certain sound means food is available while anotehr means that it isn’t.

From this point of view, neither of the two models presented in the paper is compatible with the PCT reorganization model because both assume that rats can predict (RFE model) or infer (ANCCR model) which stimulus sound is the “cause” of the availability of sucrose. As Bill said in one of his posts on condiiotning studies: “Of course we don’t assume that logical reasoning is involved in reorganization; it only looks that way”. In other words, the appearnce that logical resonng is involved in this learning task is another example of a behavioral illusion!

A reorganization model works quite differently than the RFE and ANCCR models. It doesn’t reason, it just randomly varies its output and when an output makes things better it delays longer until changing to a new output. It would be interesting to see if a reorganization model could account for the Joeng et al data as well as their ANCCR model does. Such a model would have a problem dealing with the fact that dopamine release is the dependent variable. But since the RFE and ANCCR models were compared on the basis of their ability to account for this variable I imagine that the reorganization model should be able to handle it as well. It would be nice to have another example – besides Bill’s analysis of the behavior ina shock avoidance experiment-- of PCT’s ability to handle the behavior in conditioning experiments.

Best, Rick

This is why I framed the ‘inference’ idea in terms of reorganization, saying

That’s true. We collectively have paid little attention to the huge amount of information that is available about control loops that are implemented by varying chemical concentrations. Dopamine is implicated in neural plasticity and memory consolidation. Does dopamine foster and localize reorganization?

Most of the research literature is deep into the weeds. There’s a useful overview on this commercial site. Here’s a bit focused on synaptic plasticity.

Dopamine neurotransmitters bind to five subtypes of dopamine receptors: D1, D2, D3, D4, and D5, which are members of the G-protein coupled receptor family that are divided into two major subclasses: D-1-like and D-2-like. The binding of dopamine to these receptors initiates cascades of signaling responsible for activating functions in the associated areas of the brain where each receptor type is most prevalent. D1-like receptors are more prevalent than D2-like receptors. To understand how dopamine functions in the human brain as a neurotransmitter requires looking at the effect of dopamine binding to D1-like and D2-like receptor types to exert their effects via second messenger systems. The binding of dopamine to D1-like receptors (D1 and D5) results in excitation via the opening of Na+ channels or inhibition via the opening of K+ channels. D1-like receptor stimulation induces the activation of adenylate cyclase, the enzyme that converts ATP to cAMP, thereby increasing cAMP levels leading to the disinhibition of protein kinase A (PKA) which phosphorylates downstream targets such as cAMP regulatory element binding protein (CREB). The translocation of CREB to the nucleus and CREB-dependent transcription of numerous genes is responsible for the synaptic plasticity necessary for learning and memory formation. In contrast, D-2 like receptor binding (D2, D3, and D4) lead to inhibition of the target neuron by exerting an opposite effect of inhibiting adenylate cyclase through coupling to G proteins Gi/o which decreases the production of cAMP. Whether dopamine is excitatory or inhibitory is a matter of which type of effect on a target neuron is exerted which is based on which types of receptors are on the membrane surface of the neuron and how the neuron responds to increases or decreases in cAMP concentration.

I guess I was just thrown by the title of this thread, which says the Joeng et al study is “compatible with PCT versus predictive coding”. Having now seen the article it seems no more compatibe with PCT than any other study of its type. Their non-predictive coding model (ANCCR) is no more PCT compatible than their predictve coding one (RPE). Both are really models of how a sound comes to be part of the process of controlling for the input of sucrose solution. The development of that control process probably does involve learning to control a perception that is some combination of both the sound and the sucrose solution but, as you note, PCT would suggest that that learning probably involves reorganization rather than either forward (RPE) or backward (ANCCR) causal inference.

I think we should know a lot more about how reorganization works at the behavioral level before trying to answer that question.

Using this kind of physiological analysis as a basis for understanding a behavioral phenomenon is like using the chemistry of sea water to understand buoyancy. Knowledge of neurophysiology is imporant and useful for constraining our behavioral models of control, as Powers did in the Premises chapter of B:CP. But developing accurate models at the behavioral level must come before trying to figure out how the behavioral level phenomenon is produced by the neurophysiology. And the behavioral models must be control models in order to be accuratre.

My impression is that, to the extent that research is being done to determine how the neurophysiology produces the behavioral phenomenon, the behavioral phenomenon is seen as a causal rather the a control process. So, with some notable exceptions (such as Henry Yin’s lab), perfectly good neurophysiology is used to explain the wrong kind of behavior, as is the case, I believe, with the Joeng et al study.

– Rick

Phenomena first. The interest, Rick, and the fun, is in finding the phenomena behind the veil of their theory-laden presentations and then applying our theory, PCT, to the phenomena. Reading the extant literature thus can be a way of identifying phenomena.

It may be more comfortable to treat the extant literature the way an interviewer views a stack of resumes in a tough job market, the way a referee for funding looks at a surplus of proposals, or the way an editor or reviewer for publication looks at a surplus of manuscripts, that is, looking for reasons to ignore them.

But we don’t have an analogous surplus, and on the other hand we are, or should be, concerned about physiological plausibility of the computational theory.

Or one may regard all those current publications as expressions of prior history, pre-PCT, in the same category as, say, writings about the luminiferous ether. To do this is to embrace the fantasy that PCT has already won the field(s).

But PCT has not won any field, and will not do so without identifying phenomena that are discussed in theoretical terms that we don’t like, and engaging the discussants with a PCT alternative for those same phenomena, with reference to what they have written.

This is work for younger men and women than thee and me, I think, but it’s worthwhile identifying phenomena and their current descriptions, for further investigation.

I know. That’s what I did, didn’t I? Based on what was observed I saw the observations made by Joeng et al as having something to do with how new perceptual functions are built, not on what those functions are.

Well, it’s not always that much fun. I submit for your consideration my PCT explanation of a phenomenon called the power law of movement.

The phenomena relevant to PCT pertain to the controlling done by living systems. The phenomena described in the extant behavioral science literature are usually irrelevant side effects of controllng (such as the power law). As I note in The Study of Living Control Systems, sometimes the phenomena described in this literature can provide hints about control. But such phenomena also can be very misleading.

We definitely should be concerned about the physiological plausibility of PCT. But this should follow (or occur in tandem with) tests of the PCT model against data relevant to the phenoenon that PCT was developed to explain: control (known to the layman as purposeful behavior)

This assumes that PCT is an alternative explanation of the phenomena described in most of the extant literature. It is not! It is an explanation of a phenomenon that current behavioral scientists don’t even recognize as a phenomenon. I’ll explain this in my upcoming post on why I have been putting so much work into showing that the power law of movement is an irrelevabt side effect of controlling.

I think it’s worthwhile bring to the attention of this group those papers in the extant literature that describe phenomena that are clearly relevant to controlling; where it’s pretty clear what variables are being controlled. But I don’t think it would be worth the time of anyone, no matter their age, to go searching though the extant literature. But for those who have nothing better to do then, I say go ahead and search away!

Yes, that was also my understanding.

If indeed they are constructing perceptual functions that perceive causal relationships. as it appears from the description, that’s pretty interesting. If dopamine is instrumental in making synaptic changes that bring this about, that’s actually quite interesting! A significant difficulty in the theory of reorganization is: how is it localized to just the places in the hierarchy that need to change? How does spread, when ineffective, and how is that spread throttled so as not to blossom system-wide? Local production of neurochemicals seems an obvious place to look, at first inter-synaptic, then in the intercellular environment. The war of the soups and the sparks was long ago resolved to comity, a duality not so opaque to reason as particle-wave radiation but still having its obscurities. PCT talk is stuck firmly in the sparks camp, with only vague hand-waving correspondence of the ‘soup’ phenomena to weights and fudges in our computational models.

How does that bit of the standard complaint apply to this case? They demonstrated that what developed was control of causal relationships, as distinct from a supposed processing of predictive error. It’s about reorganization creating functions for perception and control. I see no confusion of side effects here.