Gradient-Based Learning of Compositional Dynamics with Modular RNNs

Abstract

Learning compositional dynamics with recurrent neural networks (RNNs) trained with back-propagation through time (BPTT) is usually a difficult task. Typically RNNs learn the consecutive shape along target sequences from time step to time step, focusing on local temporal correlations. When the challenge is to identify and model independent, unknown data subcomponents, that is, data generating causes on-the-fly during training, however, this local temporal shape-oriented inductive learning bias is obstructive. We propose a modular, compositional RNN architecture and derive simple procedures to automatically infer the source subdynamics that generate the data. We show that the involved error signal separation can be used for both teacher forcing and model-distinct target signal provision in the compositional RNN architecture. As a result, the entire network is able to learn compositional dynamics, developing emergent, flexibly adaptable signal decompositions within the distributed modules. We demonstrate that in this way simple RNNs trained with BPTT can learn sequences that could so far only be solved effectively with reservoir computing approaches. Moreover we show that these RNNs are much more robust against signal noise when compared to traditional BPTT or reservoir computing approaches.