Overcoming catastrophic forgetting in neural networks

被引：4538

作者：

Kirkpatricka, James ^{[1
]}

Pascanu, Razvan ^{[1
]}

Rabinowitz, Neil ^{[1
]}

Veness, Joel ^{[1
]}

Desjardins, Guillaume ^{[1
]}

Rusu, Andrei A. ^{[1
]}

Milan, Kieran ^{[1
]}

Quan, John ^{[1
]}

Ramalho, Tiago ^{[1
]}

Grabska-Barwinska, Agnieszka ^{[1
]}

Hassabis, Demis ^{[1
]}

Clopath, Claudia ^{[2
]}

Kumaran, Dharshan ^{[1
]}

Hadsell, Raia ^{[1
]}

机构：

[1] DeepMind, London EC4 5TW, England

[2] Imperial Coll London, Dept Bioengn, London SW7 2AZ, England

来源：

PROCEEDINGS OF THE NATIONAL ACADEMY OF SCIENCES OF THE UNITED STATES OF AMERICA | 2017年 / 114卷 / 13期

基金：

英国工程与自然科学研究理事会; 英国惠康基金;

关键词：

synaptic consolidation; artificial intelligence; stability plasticity; continual learning; deep learning; COMPLEMENTARY LEARNING-SYSTEMS; CONNECTIONIST MODELS; PREFRONTAL CORTEX; MEMORY;

D O I：

10.1073/pnas.1611835114

中图分类号：

O [数理科学和化学]; P [天文学、地球科学]; Q [生物科学]; N [自然科学总论];

学科分类号：

07 ; 0710 ; 09 ;

摘要：

The ability to learn tasks in a sequential fashion is crucial to the development of artificial intelligence. Until now neural networks have not been capable of this and it has been widely thought that catastrophic forgetting is an inevitable feature of connectionist models. We show that it is possible to overcome this limitation and train networks that can maintain expertise on tasks that they have not experienced for a long time. Our approach remembers old tasks by selectively slowing down learning on the weights important for those tasks. We demonstrate our approach is scalable and effective by solving a set of classification tasks based on a hand-written digit dataset and by learning several Atari 2600 games sequentially.

引用

页码：3521 / 3526

页数：6

共 44 条

[31]

McCloskey M., 1989, Psychology of learning and motivation, V24, P109, DOI DOI 10.1016/S0079-7421(08)60536-8

[32]

Milan Kieran., 2016, Advances in Neural Information Processing Systems

[33] An integrative theory of prefrontal cortex function [J].