Deep Bayesian Neural Networks for Continual Learning

Abstract

In the problem of continual learning, tasks are given in sequence and the goal of the machine is to learn new tasks while retaining previously learned tasks' performances. Even though deep neural networks have become prevalent among machine learning techniques recently, they may fail to deal with continual learning environment. As the model learns new tasks, parameters contributing largely to previous tasks' performances may change and result in poor performance. This phenomenon is called catastrophic forgetting, and researchers tackled this problem using regularizers and structure optimization techniques.

We aim to show that applying Bayesian modelling to deep neural network is beneficial for continual learning. Not only does Bayesian framework provide systematic way of performing online learning, it also provides uncertainty estimates for measuring each parameter's contribution to previously learned tasks' performances. By rescaling gradients applied to parameters with large performance contribution, the model can retain performances to previous tasks longer. This thesis shows how deep Bayesian neural networks utilize model uncertainty, which leads to alleviation of catastrophic forgetting in continual learning environment.