Towards Encoding Background Knowledge with Temporal Extent into Neural Networks

Abstract

Neuro-symbolic integration merges background knowledge and neural networks to provide a more effective learning system. It uses the Core Method as a means to encode rules. However, this method has several drawbacks in dealing with rules that have temporal extent. First, it demands some interface with the world which buffers the input patterns so they can be represented all at once. This imposes a rigid limit on the duration of patterns and further suggests that all input vectors be the same length. These are troublesome in domains where one would like comparable representations for patterns that are of variable length (e.g. language). Second, it does not allow dynamic insertion of rules conveniently. Finally and also most seriously, it cannot encode rules having preconditions satisfied at non-deterministic time points – an important class of rules. This paper presents novel methods for encoding such rules, thereby improves and extends the power of the state-of-the-art neuro-symbolic integration.