Mateusz Wasiluk

Bio-inspired methods for continual learning in deep neural networks

In the past decade, the field of Artificial Intelligence (AI), and more precisely, Machine Learning (ML) has experienced enormous growth in terms of scientific interest and the number of industrial applications. While self-improving algorithms have already become widespread in the modern world, largely due to their ability to specialise, there are still scenarios in which they fail to meet the requirements of a more challenging training process. One such scenario is continual or lifelong learning, in which a model experiences a sequential stream of training episodes and needs to adjust to novel input distributions or tasks while still performing well on the previous ones. Despite such an ability being a standard property of biological cognitive systems (like humans), the lack of it in artificial neural networks (ANNs) is still an open problem. It manifests itself mainly as the phenomenon of catastrophic forgetting (CF) in which a new set of patterns suddenly and completely erases a network’s previous knowledge. Most of the current methods of overcoming this issue either strongly specialise in a particular area of applications (like image recognition) or are too computationally expensive to be used efficiently (e.g. by storing all the incoming data and regularly retraining the whole model). The main goal of this project is to draw from mechanisms that enable continual learning in biological agents and use them as an inspiration for novel methods of preventing catastrophic forgetting in artificial neural networks.