Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Continual Learning Should Move Beyond Incremental Classification

Rupert Mitchell, Antonio Alliegro, Raffaello Camoriano, Dustin Carrión-Ojeda, Antonio Carta, Georgia Chalvatzaki, Nikhil Churamani, Carlo D'Eramo, Samin Hamidi, Robin Hesse, Fabian Hinder, Roshni Ramanna Kamath, Vincenzo Lomonaco, Subarnaduti Paul, Francesca Pistilli, Tinne Tuytelaars, Gido M van de Ven, Kristian Kersting, Simone Schaub-Meyer, Martin Mundt

TL;DR

The paper contends that continual learning has been overly tethered to incremental classification and proposes moving toward a framework that explicitly handles continuity, spaces, and learning objectives. Through concrete examples, it demonstrates limits of current CL methods in multi-target regression, constrained-output robotics, continuous task spaces, and abstract memorization. It offers principled recommendations, including formalizing temporal dynamics with Distribution Processes, embracing continuous task identities, and integrating generative/density-based objectives, to broaden the scope and strengthen the theory of CL. This shift aims to produce learning systems that are theoretically sound and practically applicable to real-world problems beyond discrete classification.

Abstract

Continual learning (CL) is the sub-field of machine learning concerned with accumulating knowledge in dynamic environments. So far, CL research has mainly focused on incremental classification tasks, where models learn to classify new categories while retaining knowledge of previously learned ones. Here, we argue that maintaining such a focus limits both theoretical development and practical applicability of CL methods. Through a detailed analysis of concrete examples - including multi-target classification, robotics with constrained output spaces, learning in continuous task domains, and higher-level concept memorization - we demonstrate how current CL approaches often fail when applied beyond standard classification. We identify three fundamental challenges: (C1) the nature of continuity in learning problems, (C2) the choice of appropriate spaces and metrics for measuring similarity, and (C3) the role of learning objectives beyond classification. For each challenge, we provide specific recommendations to help move the field forward, including formalizing temporal dynamics through distribution processes, developing principled approaches for continuous task spaces, and incorporating density estimation and generative objectives. In so doing, this position paper aims to broaden the scope of CL research while strengthening its theoretical foundations, making it more applicable to real-world problems.

Continual Learning Should Move Beyond Incremental Classification

TL;DR

The paper contends that continual learning has been overly tethered to incremental classification and proposes moving toward a framework that explicitly handles continuity, spaces, and learning objectives. Through concrete examples, it demonstrates limits of current CL methods in multi-target regression, constrained-output robotics, continuous task spaces, and abstract memorization. It offers principled recommendations, including formalizing temporal dynamics with Distribution Processes, embracing continuous task identities, and integrating generative/density-based objectives, to broaden the scope and strengthen the theory of CL. This shift aims to produce learning systems that are theoretically sound and practically applicable to real-world problems beyond discrete classification.

Abstract

Continual learning (CL) is the sub-field of machine learning concerned with accumulating knowledge in dynamic environments. So far, CL research has mainly focused on incremental classification tasks, where models learn to classify new categories while retaining knowledge of previously learned ones. Here, we argue that maintaining such a focus limits both theoretical development and practical applicability of CL methods. Through a detailed analysis of concrete examples - including multi-target classification, robotics with constrained output spaces, learning in continuous task domains, and higher-level concept memorization - we demonstrate how current CL approaches often fail when applied beyond standard classification. We identify three fundamental challenges: (C1) the nature of continuity in learning problems, (C2) the choice of appropriate spaces and metrics for measuring similarity, and (C3) the role of learning objectives beyond classification. For each challenge, we provide specific recommendations to help move the field forward, including formalizing temporal dynamics through distribution processes, developing principled approaches for continuous task spaces, and incorporating density estimation and generative objectives. In so doing, this position paper aims to broaden the scope of CL research while strengthening its theoretical foundations, making it more applicable to real-world problems.

Paper Structure

This paper contains 11 sections, 4 equations, 4 figures.

Table of Contents

  1. Introduction
  2. Core Examples
  3. How well addressed is supervised CL for classification?
  4. How can CL accommodate constraints?
  5. What is a task? CL in continuous domains.
  6. What is memorable?
  7. Conceptual Framing: Where from here?
  8. On Continuity
  9. On Spaces
  10. On Objectives
  11. Concluding Remarks

Figures (4)

  • Figure 1: Map of diverse facial expressions on Arousal-Valence axes. This representation captures the inherently continuous variation of expressions as opposed to, e.g., "angry" and "sad".
  • Figure 2: Depiction of a trajectory learned from demonstrations (shown in red) on the surface of a sphere. This example illustrates the challenge of constrained structured prediction in robotics, where valid outputs must lie on a two-dimensional manifold (the sphere's surface) within a three-dimensional space. Traditional continual learning approaches using Euclidean distance metrics may fail to maintain such geometric constraints during learning.
  • Figure 3: A robot arm pushing a box onto a target marker (green). The arm makes contact at a single point and must adjust for the weight distribution in the box. Image from tiboni2024domain.
  • Figure 4: Zergling rush in Starcraft II: the blue player (with the tan/blue buildings) has failed to completely block the entrance at the lower right, allowing zerglings (small and red) into their base.