École Nationale Supérieure de
Techniques Avancées

Abstract

Neuromorphic computing aims at doing computation in a fast and energy-efficient way by taking inspiration from the brain. Neuromorphic computing involves the researches of in-memory computing architectures, computational principles and also learning algorithms. It is important for the neuromorphic system to apply a compatible learning algorithm which enables the system to realize an admiring performance with high reaction speeds and minimal power consumption. Equilibrium Propagation(EP) is an algorithm deeply studied in the neuromorphic community because it relies on the same hardware for both the training and the inference phase and offers a local learning rule. However, EP has up to now only been studied in the supervised (i.e with labelled data) learning fashion, while in practice the large amount of labelled data is not always available. Here is an attempt to train dynamical neural network with EP in the unsupervised fashion (i.e with unlabelled data). This new learning algorithm is named as unsupervised equilibrium propagation (unsupervised EP). Starting from the initial EP algorithm, we added the mechanisms of 'winner takes all’ (WTA) and refractory period (RP) to achieve the unsupervised learning on image classification tasks. Our unsupervised EP achieved the accuracy of 95.20% on Digits dataset, 84.52% on MNIST dataset.