Towards hypervector representations for learning and planning with schemas

Abstract

The Schema Mechanism is a general learning and concept building framework initially created in the 1980s by Gary Drescher. It was inspired by the constructivist theory of early human cognitive development by Jean Piaget and shares interesting properties with human learning. Recently, Schema Networks were proposed. They combine ideas of the original Schema mechanism, Relational MDPs and planning based on Factor Graph optimization. Schema Networks demonstrated interesting properties for transfer learning, i.e. the ability of zero-shot transfer. However, there are several limitations of this approach. For example, although the Schema Network, in principle, works on an object-level, the original learning and inference algorithms use individual pixels as objects. Also, all types of entities have to share the same set of attributes and the neighborhood for each learned Schema has to be of the same size. In this paper, we discuss these and other limitations of Schema Networks and propose a novel representation based on hypervectors to address some of the limitations. Hypervectors are very high dimensional vectors (e.g. 2,048 dimensional) with useful statistical properties, including high representational capacity and robustness to noise. We present a system based on a Vector Symbolic Architecture (VSA) that uses hypervectors and carefully designed operators to create representations of arbitrary objects with varying number and type of attributes. These representations can be used to encode Schemas on this set of objects in arbitrary neighborhoods. The paper includes first results demonstrating the representational capacity and robustness to noise.