7. Pattern recognition

A unitracker is basically a pattern recognition mechanism. “Unitracker” was coined, I presume, by Ruth Millikan, and developed most extensively in the first part of her book “Beyond Concepts.” The name emphasizes that a given unitracker mechanism tracks exactly one pattern, which pattern, I think, Millikan refers to as a unicept. The pattern can change over time, by adjusting the inputs and actions of the mechanism, but at any one time the mechanism recognizes one pattern. Millikan surmises that human cognition depends on (is based on?) a wide variety of such unitrackers. Millikan does not address the role of unitrackers in Consciousness, but I think this role, given the framework, is pretty straightforward.

I believe unitrackers are mechanisms whose purpose is to generate representation vehicles, so, the first of the two mechanisms involved in representation (as described on the page on representation). The “meaning” of the representation will be determined by the use to which the vehicle is put. In some cases, that use will be as input to subsequent unitrackers. For example, a neural network that recognizes cat images is not a unitracker for cats, but it is a unitracker for images of cats. A “cat” unitracker might take input from a “cat image” unitracker, as well as from a “cat sound” (meow) unitracker, as well as from a “word sound” (“there’s a cat!”) unitracker, etc. The output of a unitracker is a representation which carries mutual information with respect to the unitracker mechanism, and so too, to the target pattern (unicept) of the unitracker.

[begin wild speculation]

My current working hypothesis is that the unit of the neorcortex, as embodied by the cortical column, is the unitracker. I.e., I’m guessing that one cortical column = one unitracker. I’m guessing the entire cortex from front to back is unitrackers. Some unitrackers receive input directly from senses (via thalamus, or not?). Some unitrackers receive inputs mostly from other nearby unitrackers in a process of building more complex patterns, as in pixel -> edge -> boundary -> object -> etc.. Some (all?) unitrackers send outputs back to the input unitrackers as feedback or “prediction”. Most (all?) unitrackers send outputs to and receive non-sensory inputs from the thalamus (see next page for why) and/or other subcortical structures. Some (all?) unitrackers receive inputs from distant unitrackers, mostly from the PFC, and most of which are suppressive.

[/end wild speculation, for now]