Intrinsic representation : bootstrapping symbols from experience

Abstract

If we are to understand human-level intelligence, we need to understand how meanings can be learned without explicit instruction. I take a step toward that understanding by focusing on the symbol-grounding problem, showing how symbols can emerge from a system that looks for regularity in the experiences of its visual and proprioceptive sensory systems. More specifically, my implemented system builds descriptions up from low-level perceptual information and, without supervision, discovers regularities in that information. Then, my system, with supervision, associates the regularity with symbolic tags. Experiments conducted with the implementation shows that it successfully learns symbols corresponding to blocks in a simple 2D blocks world, and learns to associate the position of its eye with the position of its arm. In the course of this work, I take a new perspective on how to design knowledge representations, one that grapples with the internal semantics of systems, and I propose a model of an adaptive knowledge representation scheme that is intrinsic to the model and not parasitic on meanings captured in some external system, such as the head of a human investigator.