The Birth of Purpose
Biological Intelligence and Unconscious Process
What is biological intelligence? The question should be pursued not only in terms of competence, but creativity. Intelligence does not merely outperform other alternatives, it generates solutions in a richly textured unpredictable environment. The ability to generate contextually dependent solutions through the interaction of a small set of modular behaviors defined by metastabilities in statespace: ethology used to call this “instinct”, and this is where I believe we should begin.
First, a little historical overview from the perspective of a psychoanalyst:
The Freudian, Lacanian, and poststructuralist theories of unconscious process are much more compatible with mathematical modeling than generally assumed. It was Freud and Levi-Strauss who first designed a theory of symbolic motivation compatible with the vector field: existing neural network models and NLP techniques actually owe a great deal to early 20th century psycholinguistics. The inventor of the perceptron and thus the forerunner of all connectionist models, Frank Rosenblatt, was himself a psychologist. W. Ross Ashby, the first to write systematically about cybernetics in the 1950s, was also. The still fundamental “McCulloch–Pitts” neuronal model was built from the collaboration of a logician and a psychologist.
But what’s still missing is an integration between neural network principles and the rigorous formulation of unconscious mechanics: these mechanics themselves have been known to psychoanalysis for more than a century, proven in a thousand contexts, and yet remain largely unexploited in any computational context. Partly this is due to the exotic flavor continental philosophy still has for the Anglo-American sphere in which nearly all AI and neuropsychology has taken place, and partly this is due to the inevitable resistance which any psychology of the unconscious encounters: part of how unconscious process secures efficiency, is denial of its existence.
One of the insights of the mathematician Norbert Wiener, is to have conceived of a signal as a special kind of noise. The psychological analogue can help explain what I mean by “unconscious mechanics”: namely that conscious process is largely constituted by redundant marking. A conscious topology is coherent but informatively sparse. Since the Romantics of the 1800s, the realm of unconscious process as been popularly characterized as lawless, arbitrary, and irrational. Nearly the reverse is the case: what is unconscious can afford to remain so, because it’s metabolically optimized. What we’re conscious of, we do awkwardly, slowly, expensively. It was one of the important results of chaos theory in the 1990s to emphasize that chaos should not be properly considered “random”, but a state of excitation produced by the competition between many potential kinds of order, such that the total statespace of a given system lacks easily detected structure. Similarly, biological noise only appears random because the cumulative amplitude reflects too many trajectories simultaneously to be formulated as a linear function. The great biologist E.O. Wilson had a similar realization while observing his ant colonies. At low states of excitation, the collective behavior of Hymenoptera can appear futile and aimless: a single ant may pick up a pebble in the vicinity of the entrance, move it a few inches, drop it, walk in a circle, pick it up again only to move it back to its original position. But this is the stochastic foundation from which purposive behavior is built: it requires many nonlinear interactions between many relatively free agents acting within the strict restraints of a finite system, to produce holistic “agential” behavior. It was the principal insight of cybernetics that in a system’s search for homeostasis, purposive behavior is born. Biological systems are inherently dissipative, and therefore designed to withstand enormous energetic throughput such that they are capable of exploiting the fallback in generating organized purposive behavior. This kind of self-organizing novelty cannot be obtained through linear guidance, but rather must be obtained through the system’s own exploration of its critical moments.
An unconscious topology such as we experience it in psychoanalysis, is above all richly connected and overdetermined: any one effect has many causes. A hypothesis reliant on a bijective relation between intelligent result and uniquely determined mechanic will always fail. The systems we’re discussing are fully deterministic in shortterm pathfinding yet stochastic in longterm evolution - their statespace trails have rational causes but are extremely difficult to predict: my argument is that the same characterization applies to unconscious process, because although the semantic valences with which it deals are highly ordered, their extreme connectedness makes surprises the rule rather than the exception. In the classic example of displacement and the element of the uncanny in dreams, the object choice which serves to obscure the repressed content often has the appearance of the comically arbitrary, but upon closer examination was always assigned its role due to a maximal connectivity to the original repressed ideation: Freud’s phrase “Wortbrücke” was apt and prophetic, considering the current direction of natural language processing. Neurosis forms as a locus of high connective throughput, which encapsulates a region of the repressed like scar tissue. Arguably, even ancestral instinctual centers in an unconscious topology produce similar distortion effects. This is what one of the last great naturalists, the ethologist Nikolaas Tinbergen, also incidentally termed “displacement”: when the primary instinctual path is blocked, such as the consummatory act of a mating ritual, other peripheral behaviors become activated through displacement. A male goose frustrated in his courtship will instead peck at the ground furiously. In this one easily pictured example, we see both the origin for the capacity for repression - namely the halting of instinctual discharge in a social context - and the evidence of the protolinguistic relationship at work in any given displacement. Namely that what the male goose wants to do in the fulfillment of his duties is symbolically enacted in the way he pecks at the ground - a substitute behavior which is nonetheless “understood” by all.
Everything I’ve said could at least be employed in making natural language processing more interesting than it is now. In my opinion, its output is largely boring, only vaguely plausible, but not valid. It’s merely the statistical average of a great many texts organized into a vast lookup table: there is no meaning because there is no driving force, there is no driving force because there is no theory of hermeneutics behind the modeling, nor any model of a constrained system undergoing energetic stress. The wonders of GPT are obtained through bruteforce combinatorial computation: any given word’s likely occurrence in the vicinity of any given word. It’s my contention that this has nothing to do with biological intelligence nor what’s valuable about human communication: which is why its output reads like a mediocre 10th grader trying to sound smarter than they are via plagiarism and a thesaurus. We still have an opportunity of doing something more significant than improving on the latest chatbot: as a scholar and psychologist, this interests me much more, despite the fact that the big cash money wad probably lies in the direction of producing vapid results for the sake of credulous crowds. The 21st century bigtop and sideline freakshow takes place digitally, but the crushing stupidity of the majority remains the same. Yet the activity of the field tells me that I’m not the only one thinking this way, and that someone will implement these ideas in the near future…