PaperlayerView exact revision on arXiv

Source-linked AI summary

Continual Learning and Catastrophic Forgetting

Gido M. van de Ven, Nicholas Soures, Dhireesha Kudithipudi

arXiv:2403.05175v1cs.LGcs.AIcs.CVq-bio.NCstat.ML

TL;DR

Continual learning is difficult for deep networks because learning new information can trigger catastrophic forgetting, and reducing forgetting brings additional trade-offs. This chapter reviews the field’s challenges and six computational approaches, emphasizing that continual learning involves more than preventing forgetting.

  • Problem

    Deep networks tend to catastrophically forget previously learned information when trained on something new, and reducing forgetting alone does not resolve continual learning’s other challenges.

    Source: §1 · p. 2, §2.2 · p. 3

  • Method

    The chapter reviews six computational approaches: replay, parameter and functional regularization, optimization-based approaches, context-dependent processing, and template-based classification.

    Source: §5 · p. 13

  • Results

    Reducing forgetting can impair learning new information, while training on all data seen so far presents memory and computation challenges.

    Source: §2.2 · p. 3

  • Takeaways & Limitations

    Deep networks’ current inability to learn continually is a major obstacle to artificial agents accumulating knowledge from experience.

    Source: §5 · p. 13

    Context-dependent processing alone is limited to task-incremental learning when the context is assumed to be clear to the network.

    Source: §3.5 · p. 11

Abstract

from arXiv · show

This book chapter delves into the dynamics of continual learning, which is the process of incrementally learning from a non-stationary stream of data. Although continual learning is a natural skill for the human brain, it is very challenging for artificial neural networks. An important reason is that, when learning something new, these networks tend to quickly and drastically forget what they had learned before, a phenomenon known as catastrophic forgetting. Especially in the last decade, continual learning has become an extensively studied topic in deep learning. This book chapter reviews the insights that this field has generated.

Key Points

Continual learning requires accumulating knowledge from changing data streams, but artificial neural networks can catastrophically forget earlier learning. The chapter outlines the problem, evaluation dimensions, computational approaches, and links to cognitive science.

Key Points

  • Continual learning means incrementally learning from non-stationary data, while catastrophic forgetting is the rapid loss of prior learning when new information is learned.For deep neural networks, forgetting is a major challenge, but preventing it alone is not sufficient to solve continual learning.Source: Key Points · p. 1
  • The literature distinguishes task-based from task-free learning and task-, domain-, and class-incremental learning; these distinctions are orthogonal.Continual-learning evaluation covers performance, diagnostic analysis, and resource efficiency.Source: Key Points · p. 1
  • Six computational approaches are replay, parameter regularization, functional regularization, optimization-based approaches, context-dependent processing, and template-based classification.Source: Key Points · p. 1
  • Further connections between deep learning and cognitive science in continual learning could benefit both fields.Source: Key Points · p. 1

1 Introduction

Humans accumulate knowledge as they learn, while artificial deep neural networks can catastrophically forget earlier learning. This chapter reviews continual learning challenges and methods, motivated by practical costs and the opportunity for connections with cognitive science.

1 Introduction

  • When deep networks learn something new, they tend to almost completely forget what they learned before, unlike humans’ ability to integrate new information with prior knowledge.The chapter identifies this as catastrophic forgetting.Source: §1 · p. 2
  • Because updating a trained network on new data alone does not work reliably, practitioners may retrain from scratch on all data despite substantial computational costs.Continual-learning methods could bring efficiency gains and reduce required resources.Source: §1 · p. 2
  • Studying deep networks’ continual-learning failures may offer clues about brain processes, while the brain’s abilities may inspire new algorithms.Source: §1 · p. 2

2 The Continual Learning Problem

The chapter defines continual learning as accumulating knowledge from changing data streams and surveys why it is difficult, how its settings differ, and how methods are evaluated. It emphasizes that reducing forgetting is only one part of the problem.

2 The Continual Learning Problem

  • Continual learning is incremental learning from a non-stationary data stream, where the data distribution changes over time and new learning should accumulate rather than overwrite prior knowledge.Source: §2 · p. 2

2.1 Catastrophic Forgetting

  • Catastrophic forgetting is rapid, drastic loss of earlier learning during new learning; sequential training can cause it even after small amounts of training on new data.Interleaved training lets the same network learn both tasks, indicating that forgetting is not due to limited model capacity.Source: §2.1 · pp. 2–3, Figure 1 · p. 3

2.2 Other Features Important for Continual Learning

  • Reducing forgetting alone is insufficient: it can impair learning new information, while training on all past data raises memory and computation challenges.Other desired capabilities include rapid adaptation, positive transfer, task agnosticism, noise tolerance, and resource efficiency.Source: §2.2 · pp. 3–4

2.3 Task-based versus Task-free Continual Learning

  • Task-based learning presents discrete tasks sequentially, often with marked boundaries; task-free learning permits gradual transitions and changing task probabilities.Task-based settings offer controlled study, while hard-boundary reliance and repeated tasks motivate task-free settings.Source: §2.3 · pp. 4–5, Figure 2 · p. 4
  • The task set specifies what changes in the data stream, while the schedule specifies how it changes over time; continuous task sets remain largely unexplored.Source: §2.3 · p. 5

2.4 Three Continual Learning Scenarios

  • Task-incremental learning assumes task identity is known; the task-, domain-, and class-incremental scenarios are orthogonal to task-based versus task-free training.In class-incremental learning, the network must distinguish classes not observed together; domain-incremental learning changes input distributions without necessarily identifying the domain.Source: §2.4 · pp. 5–6, Figure 3 · p. 6

2.5 Evaluation

  • Continual-learning evaluation spans performance, diagnostic analysis, and resource efficiency, including learning accuracy, transfer, computational overhead, and model or memory growth.Average performance can hide different learning dynamics, while computation may also increase during inference.Source: §2.5 · pp. 6–7

3 Continual Learning Approaches

This section reviews computational strategies rather than individual methods; the strategies target replayed data, loss terms, optimization, network routing, or classification.

3.1 Replay

  • Replay supplements current data with stored or generated samples to approximate interleaved learning, but can be costly and depends on representative samples or adequate generative quality.Replay may require only relatively small amounts of past data; selecting what to replay, how to store it, and how to combine it with other methods remain open questions.Source: §3.1 · pp. 7–8

3.2 Parameter Regularization

  • Parameter regularization penalizes changes to parameters estimated as important for prior tasks, but importance estimation is difficult and these methods often struggle with class-incremental task boundaries.Importance can be estimated with Fisher Information or online methods; Fisher-based estimation can be costly.Source: §3.2 · pp. 8–9

3.3 Functional Regularization

  • Functional regularization constrains the network’s mapping at anchor inputs; representative anchors matter, but their optimal selection remains open.It can be viewed as replay in which the network’s prior predictions are stored internally, while past inputs are stored externally.Source: §3.3 · pp. 9–10

3.4 Optimization-based Approaches

  • Optimization-based methods alter updates rather than the loss, using adaptive rates, gradient projection, or synaptic consolidation; routines finding wider minima tend to perform better in continual learning.Wider minima require larger parameter changes to escape.Source: §3.4 · pp. 10–11

3.5 Context-dependent Processing

  • Context-dependent processing uses task-specific network parts, gates, or added components to reduce interference, but requires clear context or a separate context-identification method beyond task-incremental settings.Using a separate network per task eliminates interference but also removes the possibility of positive transfer.Source: §3.5 · p. 11

3.6 Template-based Classification

  • Template-based classification learns a template for each class and compares templates at test time; prototypes can reduce storage with a frozen embedding network, while generative models avoid storing samples but can make inference costly.When an embedding network must be updated, stored examples or drift correction can address prototype drift.Source: §3.6 · pp. 11–12

4 Continual Learning in Deep Learning versus in Cognitive Science

The chapter contrasts deep learning’s aim of engineering continual learners with cognitive science’s aim of explaining the brain’s existing ability. It connects biological forgetting to neural-network forgetting and describes human and artificial learning comparisons.

  • Deep learning aims to engineer artificial neural networks that learn continually, whereas cognitive science seeks to understand how the brain implements this ability.Source: §4 · p. 12
  • Cognitive science distinguishes passive and active forms of forgetting; retroactive interference from new memories is the biological form most akin to continual-network forgetting, and biological mechanisms have inspired neural approaches.Source: §4 · p. 12
  • Human and artificial-network studies found opposite effects of blocked training, while added mechanisms made network learning dynamics more human-like.Source: §4 · p. 12

5 Conclusion

Deep networks’ inability to learn continually obstructs agents’ accumulation of knowledge from experience. A complete continual-learning solution must address more than forgetting, across varied capabilities, settings, and evaluation metrics.

5 Conclusion

  • Deep networks’ current inability to learn continually is a major obstacle to artificial agents accumulating knowledge through experience.The chapter notes that deep learning models currently lack the ability to continually learn from incoming data.Source: §5 · p. 13
  • A complete continual-learning solution must address adaptation, task similarity, task-agnostic operation, noise tolerance, and efficient, sustainable resource use—not forgetting alone.The chapter identifies catastrophic forgetting as the best-known issue but says it is not sufficient to solve the problem.Source: §5 · p. 13
  • Continual learning challenges vary substantially by setup, so evaluation needs varied settings and benchmarks as well as performance, diagnostic, and resource-efficiency metrics.Examples include task-, domain-, and class-incremental learning, and task-based and task-free learning.Source: §5 · p. 13
  • The chapter reviews six approaches: replay, parameter and functional regularization, optimization-based methods, context-dependent processing, and template-based classification.The approaches modify loss functions, optimization, computation distribution, or classification, respectively.Source: §5 · p. 13
  • Interdisciplinary work drawing on neuroscience, cognitive science, and psychology may inform continual learning in artificial systems, while deep-learning insights may help explain continual learning in brains.Source: §5 · p. 13
Loading 2403.05175v1…