The Real Problem with ‘Correlation’

by rsbakker

stick zombies

Since presuming that intentional cognition can get behind intentional cognition belongs to the correlation problem, any attempt to understand the problem requires we eschew theoretical applications of intentional idioms. Getting a clear view, in other words, requires that we ‘zombify’ human cognition, adopt a thoroughly mechanical vantage that simply ignores intentionality and intentional properties. As it so happens, this is the view that commands whatever consensus one can find regarding these issues. Though the story I’ll tell is a complicated one, it should also be a noncontroversial one, at least insofar as it appeals to nothing more than naturalistic platitudes.

I first started giving these ‘zombie interpretations’ of different issues in philosophy and cognitive science a few years back.[1] Everyone in cognitive science agrees that consciousness and cognition turn on the physical somehow. This means that purely mechanical descriptions of the activities typically communicated via intentional idioms have to be relevant somehow (so long as they are accurate, at least). The idea behind ‘zombie interpretation’ is to explain as much as possible using only the mechanistic assumptions of the biological sciences—to see how far generalizing over physical processes can take our perennial attempt to understand meaning.

Zombies are ultimately only a conceit here, a way for the reader to keep the ‘explanatory gap’ clearly in view. In the institutional literature, ‘p-zombies’ are used for a variety of purposes, most famously to anchor arguments against physicalism. If a complete physical description of the world need not include consciousness, then the brute fact of consciousness implies that physicalism is incomplete. However, since this argument itself turns on the correlation problem, it will not concern us here. The point, oddly enough, is to adhere to an explanatory domain where we all pretty much agree, to speculate using only facts and assumptions belonging to the biological sciences—the idea being, of course, that these facts and assumptions are ultimately all that’s required. Zombies allow us to do that.

Philosophy Now zombie pic

So then, devoid of intentionality, zombies lurch through life possessing only contingent, physical comportments to their environment. Far from warehousing ‘representations’ possessing inexplicable intentional properties, their brains are filled with systems that dynamically interact with their world, devices designed to isolate select signals from environmental noise. Zombies do not so much ‘represent their world’ as possess statistically reliable behavioural sensitivities to their environments.

So where ‘subjects’ possess famously inexplicable semantic relations to the world, zombies possess only contingent, empirically tractable relations to the world. Thanks to evolution and learning, they just happen to be constituted such that, when placed in certain environments, gene conserving behaviours tend to reliably happen. Where subjects are thought to be ‘agents,’ perennially upstream sources of efficacy, zombies are components, subsystems at once upstream and downstream the superordinate machinery of nature. They are astounding subsystems to be sure, but they are subsystems all the same, just more nature—machinery.

What makes them astounding lies in the way their neurobiological complexity leverages behaviour out of sensitivity. Zombies do not possess distributed bits imbued with the occult property of aboutness; they do not model or represent their worlds in any intentional sense. Rather, their constitution lets ongoing environmental contact tune their relationship to subsequent environments, gradually accumulating the covariant complexities required to drive effective zombie behaviour. Nothing more is required. Rather than possessing ‘action enabling knowledge,’ zombies possess behaviour enabling information, where ‘information’ is understood in the bald sense of systematic differences making systematic differences.

A ‘cognitive comportment,’ as I’ll use it here, refers to any complex of neural sensitivities subserving instances of zombie behaviour. It comes in at least two distinct flavours: causal comportments, where neurobiology is tuned to what generally makes what happen, and correlative comportments, where zombie neurobiology is tuned to what generally accompanies what happens. Both systems allow our zombies to predict and systematically engage their environments, but they differ in a number of crucial respects. To understand these differences we need some way of understanding what positions zombies upstream their environments–or what leverages happy zombie outcomes.

The zombie brain, much like the human brain, confronts a dilemma. Since all perceptual information consists of sensitivity to selective effects (photons striking the eye, vibrations the ear, etc.), the brain needs some way of isolating the relevant causes of those effects (a rushing tiger, say) to generate the appropriate behavioural response (trip your mother-in-law, then run). The problem, however, is that these effects are ambiguous: a great many causes could be responsible. The brain is confronted with a version of the inverse problem, what I will call the medial inverse problem for reasons that will soon be clear. Since it has nothing to go on but more effects, which are themselves ambiguous, how could it hope to isolate the causes it needs to survive?

By allowing sensitivities to discrepancies between the patterns initially cued and subsequent sensory effects to select—and ultimately shape—the patterns subsequently cued. As it turns out, zombie brains are Bayesian brains.[2] Allowing discrepancies to both drive and sculpt the pattern-matching process automatically optimizes the process, allowing the system to bootstrap wide-ranging behavioural sensitivities to environments in turn. In the intentionality laden idiom of theoretical neuroscience, the brain is a ‘prediction error minimization’ machine, continually testing occurrent signals against ‘guesses’ (priors) triggered by earlier signals. Success (discrepancy minimization) quite automatically begets success, allowing the system to continually improve its capacity to make predictions—and here’s the important thing—using only sensory signals.[3]

But isolating the entities/behaviour causing sensory effects is one thing; isolating the entities/behaviour causing those entities/behaviour is quite another. And it’s here that the chasm between causal cognition and correlative cognition yawns wide. Once our brain’s discrepancy minimization processes isolate the relevant entities/behaviours—solve the medial inverse problem—the problem of prediction simply arises anew. It’s not enough to recognize avalanches as avalanches or tigers as tigers, we have to figure out what they will do. The brain, in effect, faces a second species of inverse problem, what might be called the lateral inverse problem. And once again, it’s forced to rely on sensitivities to patterns (to trigger predictions to test against subsequent signals, and so on).[4]

Nature, of course, abounds with patterns. So the problem is one of tuning a Bayesian subsystem like the zombie brain to the patterns (such as ‘avalanche behaviour’ or ‘tiger behaviour’) it needs to engage its environments given only sensory effects. The zombie brain, in other words, needs to wring behavioural sensitivities to distal processes out of a sensitivity to proximal effects. Though they are adept at comporting themselves to what causes their sensory effects (to solving the medial inverse problem), our zombies are almost entirely insensitive to the causes behind those causes. The etiological ambiguity behind the medial inverse problem pales in comparison to the etiological ambiguity comprising the lateral inverse problem, simply because sensory effects are directly correlated to the former, and only indirectly correlated to the latter. Given the limitations of zombie cognition, in other words, zombie environments are ‘black box’ environments, effectively impenetrable to causal cognition.

Part of the problem is that zombies lack any ready means of distinguishing causality from correlation on the basis of sensory information alone. Not only are sensory effects ambiguous between causes, they are ambiguous between causes and correlations as well. Cause cannot be directly perceived. A broader, engineered signal and greater resources are required to cognize its machinations with any reliability—only zombie science can furnish zombies with ‘white box’ environments. Fortunately for their prescientific ancestors, evolution only required that zombies solve the lateral inverse problem so far. Mere correlations, despite burying the underlying signal, remain systematically linked to that signal, allowing for a quite different way of minimizing discrepancies.

Zombies, once again, are subsystems whose downstream ‘componency’ consists in sensitivities to select information. The amount of environmental signal that can be filtered from that information depends on the capacity of the brain. Now any kind of differential sensitivity to an environment serves organisms in good stead. To advert to the famous example, frogs don’t need the merest comportment to fly mechanics to catch flies. All they require is a select comportment to select information reliably related to flies and fly behaviour, not to what constitutes flies and fly behaviour. And if a frog did need as much, then it would have evolved to eat something other than flies. Simple, systematic relationships are not only all that is required to solve a great number of biological problems, they are very often the only way those problems can be solved, given evolutionary exigencies. This is especially the case with complicated systems such as those comprising life.

So zombies, for instance, have no way of causally cognizing other zombies. They likewise have no way of causally cognizing themselves, at least absent the broader signal and greater computational resources provided by zombie science. As a result, they possess at best correlative comportments both to each other and to themselves.

Idoits guide to zombies

So what does this mean? What does it mean to solve systems on basis of inexpensive correlative comportments as opposed to far more expensive causal comportments? And more specifically, what does it mean to be limited to extreme versions of such comportments when it comes to zombie social cognition and metacognition?

In answer to the first question, at least three, interrelated differences can be isolated:

Unlike causal (white box) comportments, correlative (black box) comportments are idiosyncratic. As we saw above, any number of behaviourally relevant patterns can be extracted from sensory signals. How a particular problem is solved depends on evolutionary and learning contingencies. Causal comportments, on the other hand, involve behavioural sensitivity to the driving environmental mechanics. They turn on sensitivities to upstream systems that are quite independent of the signal and its idiosyncrasies.

Unlike causal (white box) comportments, correlative (black box) comportments are parasitic, or differentially mediated. To say that correlative comportments are ‘parasitic’ is to say they depend upon occluded differential relations between the patterns extracted from sensory effects and the environmental mechanics they ultimately solve. Frogs, once again, need only a systematic sensory relation to fly behaviour, not fly mechanics, which they can neglect, even though fly mechanics drives fly behaviour. A ‘black box solution’ serves. The patterns available in the sensory effects of fly behaviour are sufficient for fly catching given the cognitive resources possessed by frogs. Correlative comportments amount to the use of ‘surface features’—sensory effects—to anticipate outcomes driven by otherwise hidden mechanisms. Causal comportments, which consist of behavioural sensitivities (also derived from sensory effects) to the actual mechanics involved, are not parasitic in this sense.

Unlike causal (white box) comportments, correlative (black box) comportments are ecological, or problem relative. Both causal comportments and correlative comportments are ‘ecological’ insofar as both generate solutions on the basis of finite information and computational capacity. But where causal comportments solve the lateral inverse problem via genuine behavioural sensitivities to the mechanics of their environments, correlative comportments (such as that belonging to our frog) solve it via behavioural sensitivities to patterns differentially related to the mechanics of their environments. Correlative comportments, as we have seen, are idiosyncratically parasitic upon the mechanics of their environments. The space of possible solutions belonging to any correlative comportment is therefore relative to the particular patterns seized upon, and their differential relationships to the actual mechanics responsible. Different patterns possessing different systematic relationships will possess different ‘problem ecologies,’ which is to say, different domains of efficacy. Since correlative comportments are themselves causal, however, causal comportments apply to all correlative domains. Thus the manifest ‘objectivity’ of causal cognition relative to the ‘subjectivity’ of correlative cognition. 

So far, so good. Correlative comportments are idiosyncratic, parasitic, and ecological in a way that causal comportments are not. In each case, what distinguishes causal comportments is an actual behavioural sensitivity to the actual mechanics of the system. Zombies are immersed in potential signals, awash in causal differences, information, that could make a reproductive difference. The difficulties attendant upon the medial and lateral inverse problems, the problems of what and what-next, render the extraction of causal signals enormously difficult, even when the systems involved are simple. The systematic nature of their environments, however, allow them to use behavioural sensitivities as ‘cues,’ signals differentially related to various systems, to behaviourally interact with those systems despite the lack of any behavioural sensitivity to their particulars. So in research on contingencies, for instance, the dependency of ‘contingency inferences’ on ‘sampling,’ the kinds of stimulus input available, has long been known, as have the kinds of biases and fallacies that result. Only recently, however, have researchers realized the difficulty of accurately making such inferences given the kinds of information available in vivo, and the degree to which we out and out depend on so-called ‘pseudocontingency heuristics’ [5]. Likewise, research into ‘spontaneous explanation’ and  ‘essentialism,’ the default attribution of intrinsic traits and capacities in everyday explanation, clearly suggests that low-dimensional opportunism is the rule when it comes to human cognition.[6] The more we learn about human cognition, in other words, the more obvious the above story becomes.

So then what is the real problem with correlation? The difficulty turns on the fact that black box cognition, solving systems via correlative cues, can itself only be cognized in black box terms.

Given their complexity, zombies are black boxes to themselves as much to others. And this is what has cued so much pain behaviour in so many zombie philosophers. As a black box, zombies cannot cognize themselves as black boxes: the correlative nature of their correlative comportments utterly escapes them (short, once again, the information provided by zombie science). Zombie metacognition is blind to the structure and dynamics of zombie metacognition, and thus prone to what might be called ‘white box illusions.’ Absent behavioural sensitivity to the especially constrained nature of their correlative comportments to themselves, insufficient data is processed in the same manner as sufficient data, thus delivering the system to ‘crash space,’ domains rendered intractable by the systematic misapplication of tools adapted to different problem ecologies. Unable to place themselves downstream their incapacity, they behave as though no such incapacity exists, suffering what amounts to a form of zombie anosognosia.

Perhaps this difficulty shouldn’t be considered all that surprising: after all, the story told here is a white box story, a causal one, and therefore one requiring extraction from the ambiguities of effects and correlations. The absence of this information effectively ‘black-boxes’ the black box nature of correlative cognition. Zombies cued to solve for that efficacy accordingly run afoul the problem of processing woefully scant data as sufficient, black boxes as white boxes, thus precluding the development of effective, behavioural sensitivities to the actual processes involved.  The real Problem of Correlation, in other words, is that correlative modes systematically confound cognition of correlative comportments. Questions regarding the nature of our correlative comportments simply do not lie within the problem space of our correlative comportments—and how could they, when they’re designed to solve absent sensitivity to what’s actually going on?

And this is why zombies not only have philosophers, they have a history of philosophy as well. White box illusions have proven especially persistent, despite the spectacular absence of systematic one-to-one correspondences between the apparent white box that zombies are disposed to report as ‘mind’ and the biological white box emerging out of zombie science. Short any genuine behavioural sensitivity to the causal structure of their correlative comportments, zombies can at most generate faux-solutions, reports anchored to the systematic nature of their conundrum, and nothing more. Like automatons, they endlessly report low-dimensional, black box posits the way they report high-dimensional environmental features—and here’s the thing—using the very same terms that humans use. Zombies constantly utter terms like ‘minds,’ ‘experiences,’ ‘norms,’ and so on. Zombies, you could say, possess a profound disposition to identify themselves and each other as humans.

Just like us.

MJ zombie

 

Notes

[1] See, Davidson’s Fork: An Eliminativist Radicalization of Radical Interpretation, The Blind Mechanic, The Blind Mechanic II: Reza Negarestani and the Labour of Ghosts, Zombie Interpretation: Eliminating Kriegel’s Asymmetry Argument, and Zombie Mary versus Zombie God and Jesus: Against Lawrence Bonjour’s “Against Materialism”

[2] For an overview of Bayesian approaches, see Andy Clark, “Whatever next? Predictive brains, situated agents, and the future of cognitive science.”

[3]  The following presumes an ecological (as opposed to an inferential) understanding of the Bayesian brain. See Nico Orlandi, “Bayesian perception is ecological perception.”

[4] Absent identification there is no possibility of prediction. The analogy between this distinction and the ancient distinction between being and becoming (or even the modern one between the transcendental and the empirical) is interesting to say the least.

[5] See Klaus Fiedler et al, “Pseudocontingencies: Logically Unwarranted but Smart Inferences.”

[6] See Andrei Cimpian, “The Inherence Heuristic: Generating Everyday Explanations,” or Cimpian and Salomon, “The inherence heuristic: An intuitive means of making sense of the world, and a potential precursor to psychological essentialism.”