Efficient continual learning in neural networks with embedding regularization

Abstract

Continual learning of deep neural networks is a key requirement for scaling them up to more complex applicative scenarios and for achieving real lifelong learning of these architectures. Previous approaches to the problem have considered either the progressive increase in the size of the networks, or have tried to regularize the network behavior to equalize it with respect to previously observed tasks. In the latter case, it is essential to understand what type of information best represents this past behavior. Common techniques include regularizing the past outputs, gradients, or individual weights. In this work, we propose a new, relatively simple and efficient method to perform continual learning by regularizing instead the network internal embeddings. To make the approach scalable, we also propose a dynamic sampling strategy to reduce the memory footprint of the required external storage. We show that our method performs favorably with respect to state-of-the-art approaches in * Corresponding author. Phone: +39 06 44585495.the literature, while requiring significantly less space in memory and computational time. In addition, inspired inspired by to recent works, we evaluate the impact of selecting a more flexible model for the activation functions inside the network, evaluating the impact of catastrophic forgetting on the activation functions themselves.

Professionalization of Online Gaming? Theoretical and Empirical Analysis for a Monopoly-Holding Platform

“…In a recent study, the authors of [38] developed a deep learning framework to ensure a more realistic learning analysis. According to the authors, CL is a new, simple, and efficient method proven to be valid as an alternative to standard regularization techniques.…”

“…While Ref. [38] adopted CL in the context of neural networks, we used this method in the broad context of machine learning. The idea of CL is to allow a refined treatment of covariates based on a bias-variance trade-off argument since it consists of a two-step approach that performs a bias-variance decomposition.…”

State-of-the-Art Techniques in Artificial Intelligence for Continual Learning: A Review

“…Usually, rehearsal and progressive strategies, performs very well but always declines as the number of tasks increase, and might require a high computational power. With some differences from the first two approaches, the implementation of regularization strategy is quite simple, they require little memory, but it performance might not be up to that of rehearsal methods [7]. One main problem encountered when applying regularization strategy is determining what task best represents the behaviour of the network and, this can lead to the form of regularization penalty that would be taken [7].…”

“…With remarkable successes accomplish over the past few years in AI, deep network applications are however restricted to sole, distinct problem. Where every single network has to be trained and re-trained from the beginning every single time a new task is fed into the network and as a result their training remains very challenging to deal with particularly in real-world settings and in situations where data are scarce and/or computation is costly [7]. Furthermore, the sequence of tasks may not be clearly labelled tasks and they may switch randomly, leading to an individual task recuring in long time intervals.…”

“…Most of the algorithms presented in this paper are based on the current state and advancements in both neurophysiology and computational neuroscience field that are capable of continual learning in AI. [7,9] In this paper, we review some of the major works in continual learning both in advanced animals such as humans, whales, dogs, dolphins etc. and AI agent: we focus on how humans and animals acquire new knowledge and memories and at the same time been able to retain the useful ones over time.…”