REMEMBERING FOR THE RIGHT REASONS: EXPLANATIONS REDUCE CATASTROPHIC FORGETTING

Abstract

The goal of continual learning (CL) is to learn a sequence of tasks without suffering from the phenomenon of catastrophic forgetting. Previous work has shown that leveraging memory in the form of a replay buffer can reduce performance degradation on prior tasks. We hypothesize that forgetting can be further reduced when the model is encouraged to remember the evidence for previously made decisions. As a first step towards exploring this hypothesis, we propose a simple novel training paradigm, called Remembering for the Right Reasons (RRR), that additionally stores visual model explanations for each example in the buffer and ensures the model has "the right reasons" for its predictions by encouraging its explanations to remain consistent with those used to make decisions at training time. Without this constraint, there is a drift in explanations and increase in forgetting as conventional continual learning algorithms learn new tasks. We demonstrate how RRR can be easily added to any memory or regularizationbased approach and results in reduced forgetting, and more importantly, improved model explanations. We have evaluated our approach in the standard and few-shot settings and observed a consistent improvement across various CL approaches using different architectures and techniques to generate model explanations and demonstrated our approach showing a promising connection between explainability and continual learning. Our code is available at https://github.com/ SaynaEbrahimi/Remembering-for-the-Right-Reasons.

1. INTRODUCTION

Humans are capable of continuously learning novel tasks by leveraging their lifetime knowledge and expanding them when they encounter a new experience. They can remember the majority of their prior knowledge despite the never-ending nature of their learning process by simply keeping a running tally of the observations distributed over time or presented in summary form. The field of continual learning or lifelong learning (Thrun & Mitchell, 1995; Silver et al., 2013) aims at maintaining previous performance and avoiding so-called catastrophic forgetting of previous experience (McCloskey & Cohen, 1989; McClelland et al., 1995) when learning new skills. The goal is to develop algorithms that continually update or add parameters to accommodate an online stream of data over time. An active line of research in continual learning explores the effectiveness of using small memory budgets to store data points from the training set (Castro et al., 2018; Rajasegaran et al., 2020; Rebuffi et al., 2017; Wu et al., 2019) , gradients (Lopez-Paz et al., 2017) , or storing an online generative model that can fake them later (Shin et al., 2017) . Memory has been also exploited in the form of accommodating space for architecture growth and storage to fully recover the old performance when needed (Ebrahimi et al., 2020b; Rusu et al., 2016) . Some methods store an old snapshot of the model to distill the features (Li & Hoiem, 2016) or attention maps (Dhar et al., 2019) between the teacher and student models.

