Bidirectional Learning in Recurrent Neural Networks Using Equilibrium Propagation

Abstract

Neurobiologically-plausible learning algorithms for recurrent neural networks that can perform supervised learning are a neglected area of study. Equilibrium propagation is a recent synthesis of several ideas in biological and artificial neural network research that uses a continuous-time, energy-based neural model with a local learning rule. However, despite dealing with recurrent networks, equilibrium propagation has only been applied to discriminative categorization tasks. This thesis generalizes equilibrium propagation to bidirectional learning with asymmetric weights. Simultaneously learning the discriminative as well as generative transformations for a set of data points and their corresponding category labels, bidirectional equilibrium propagation utilizes recurrence and weight asymmetry to share related but non-identical representations within the network. Experiments on an artificial dataset demonstrate the ability to learn both transformations, as well as the ability for asymmetric-weight networks to generalize their discriminative training to the untrained generative task.