Epistemic constraints on autonomous symbolic representation in natural and artificial agents

We set out to address, in the form of a survey, the fundamental constraints upon self-updating representation in cognitive agents of natural and artificial origin. The foundational epistemic problem encountered by such agents is that of distinguishing errors of representation from inappropriateness of the representational framework. Resolving this conceptual difficulty involves ensuring the empirical falsifiability of both the representational hypotheses and the entities so represented, while at the same time retaining their epistemic distinguishability. We shall thus argue that perception-action frameworks provide an appropriate basis for the development of an empirically meaningful criterion for validating perceptual categories. In this scenario, hypotheses about the agent’s world are defined in terms of environmental affordances (characterised in terms of the agent’s active capabilities). Agents with the capability to hierarchically-abstract this framework to a level consonant with performing syntactic manipulations and making deductive conjectures are consequently able to form an implicitly symbolic representation of the environment within which new, higher-level, modes of environment manipulation are implied (e.g. tool-use). This abstraction process is inherently open-ended, admitting a wide-range of possible representational hypotheses — only the form of the lowest-level of the hierarchy need be constrained a priori (being the minimally sufficient condition necessary for retention of the ability to falsify high-level hypotheses). In biological agents capable of autonomous cognitive-updating, we argue that the grounding of such a priori ‘bootstrap’ representational hypotheses is ensured via the process of natural selection