Apperception

Suppose you have received a sequence of sensory readings and wish to “make sense” of that sequence. The standard approach to this situation is to treat it as a supervised learning problem, where we seek to maximise the conditional probability of the sensor readings at one time given readings at earlier times. In this talk, I argue that this standard approach is inadequate in various ways, and propose an alternative approach based on Kant’s notion of the “synthetic unity of apperception”. I formalise this alternative, and describe a computer implementation of Kant’s model of apperception. In a range of experiments, we found our system able to achieve satisfying interpretations of the sensory input. In particular, in multi-modal experiments, the interpretations are shown to solve the binding problem (where an object is posited that explains how different readings from different sensors can be seen as different effects of a single unified object). In occlusion experiments (where objects are sometimes obscured from view), the interpretations are shown to satisfy object-permanence. In IQ tests, our system is able to achieve human-level performance on a range of sequence prediction tasks. We stress that this is not a bespoke system hand-engineered to solve these IQ tasks; rather, it is a general system for making sense of sensory input that can, as a special case, achieve human-level performance on these IQ tasks without hand-coded domain-specific knowledge. Based on these results, I believe this formal notion of apperception to be a key component in any general intelligence.