An Integrated Nonparametric Bayesian and Deep Neural Network Framework for Biologically-Inspired Lifelong Learning

Abstract

Deep learning, trained primarily on a single task under the assumption of independent and identically distributed (i.i.d.) data, has made enormous progress in recent years. However, when naively trained sequentially on multiple tasks, without revisiting previous tasks, neural networks are known to suffer catastrophic forgetting (McCloskey and Cohen, 1989; Ratcliff, 1990), namely, the ability to perform oldtasks is often lost while learning new ones. In contrast, biological life is capable of learning many tasks throughout a lifetime from decidedly non-i.i.d. experiences, acquiring new skills and reusing old ones to learn fresh abilities, all while retaining important previous knowledge. As we strive to make artificial systems increasingly more intelligent, natural life's ability to learn continually is an important capability to emulate. Continual learning (Parisi et al., 2019) has attracted considerable attention recently in machine learning research, and a number of desiderata have emerged. Models should be able to learn multiple tasks sequentially, with the eventual number and complexity of tasks unknown. Importantly, new tasks should be learned without catastrophically forgetting previous ones, ideally without having to keep any data from previous tasks to re-train on. Models should also be capable of positive transfer: previously learned tasks should help with the learning of new tasks. Knowledge transfer between tasks maximizes sample efficiency, with this particularly important when data are scarce.

Open PDF

Document Details

Document Type
Technical Report
Publication Date
Sep 01, 2022
Accession Number
AD1180923

Entities

People

  • Aaron Courvill
  • Bing Xu
  • Christopher Kanan
  • Christopher Summerfield
  • David Warde-farley
  • Dennis Hassabis
  • Dharshan
  • Dharshan Kumaran
  • German Parisi
  • Handong Zhao
  • Ian Goodfellow
  • Ishita Dasgupta
  • Jean Pouget-abadie
  • Jianqiao Li
  • Jonathan H Cohen
  • Jose Part
  • Junya Chen
  • Kevin Liang
  • Lakshi Varshney
  • Lawrence Carin
  • Matthew Botvinick
  • Mehdi Mirza
  • Miaoyun Zhao
  • Michael Mccloskey
  • Mostafa El-ehamy
  • Nathaniel Daw
  • Neal J. Cohen
  • Nikhil Mehtra
  • Piyush Rai
  • Pk Srijith
  • Ricardo Henao
  • Roger Ratcliff
  • Ronald Kemker
  • Rui Wang
  • Ruiyi Zhang
  • Sherjil Ozair
  • Sijia Wang
  • Sreejan Kumar
  • Stefan Wermter
  • Subrata Mitra
  • Sungchul Kim
  • Thomas L. Griffiths
  • Tong Yu
  • Vinay Verma
  • Yoojin Choi
  • Yoshua Bengio
  • Yulai Cong
  • Zoubin Ghahramani

Organizations

  • Duke University

Tags

Communities of Interest

  • Autonomy
  • Energy and Power Technologies

DTIC Thesaurus Topics

  • Air Force
  • Artificial Intelligence
  • Artificial Intelligence Software
  • Computational Science
  • Computer Languages
  • Computer Programs
  • Deep Learning
  • Detection
  • Government Procurement
  • Governments
  • Information Exchange
  • Information Processing
  • Information Science
  • Information Systems
  • Language
  • Machine Learning
  • Named Entity Recognition
  • Natural Language Processing
  • Neural Networks
  • Recognition
  • Standards

Fields of Study

  • Computer science

Readers

  • Educational Psychology
  • Neural Network Machine Learning.

Technology Areas

  • AI & ML
  • AI & ML - Neural Networks