acceptodds
Under review as a conference paper at ICLR 2027

PMF-CL: Pareto-Minimal-Forgetting Continual Learner for Conflicting Tasks

Abstract

In the literature, many *continual learning* (CL) algorithms have been proposed to address the issue of *catastrophic forgetting* in ML models (i.e., learning new tasks leads to the loss of performance on previously learned tasks). Although all CL approaches use some form of memory to retain information about past tasks, a grounded understanding of what information needs to be stored to minimize catastrophic forgetting remains elusive. Recently, it has been recognized that under the strong assumption of the existence of a common global minimizer over all tasks, catastrophic forgetting can be completely avoided. However, in practice, tasks rarely have a common global minimizer, and a certain amount of forgetting is inevitable. In this paper, we propose a foundational reframing of CL as balancing conflicting tasks in hindsight from a *multi-task learning* (MTL) perspective. The approach is based on finding *Pareto-optimal* solutions, i.e., the solutions which, by definition, minimally forget the previous tasks in the Pareto sense. We derive a Pareto-minimal-forgetting CL (PMF-CL) algorithm for linear and basis-function regression. Our algorithm naturally extends to loss functions with quadratic upper-bounds around their minimizers, despite consuming only a *static* memory footprint of for model parameters, *independent* of the number of sequentially occurring tasks . In the quadratic-loss setting, we obtain exact hindsight Pareto-optimal guarantees on forgetting. In the quadratic-upper-bound setting, forgetting is provably bounded, with the tightness of the forgetting bound being directly determined by the tightness of the quadratic bound on the loss function. Our numerical results validate our theoretical claims of convergence to the Pareto-optimal solution, and demonstrate that we consistently outperform prior empirical methods while occupying lesser memory.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
continual learning, multi-task optimization, task balancing, minimal catastrophic forgetting, sequential learning

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_EhnLOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2(+1)
Soundness
2
Read review
Paper summary

The paper formulates continual learning as sequential multi-task optimization. It stores information about past losses to balance conflicting tasks, with theoretical results for quadratic losses and quadratic upper bounds.

Critical strengths
  1. The paper allows tasks to have different minimizers and directly addresses the resulting trade-offs. This is a more realistic setting than assuming all tasks share a common optimum.
  2. The method specifies what historical information needs to be retained. For linear regression, its recursive updates recover the joint-training solution without storing past samples, and the storage requirement does not grow with the number of tasks.
Critical weaknesses
  1. The experimental comparison uses different resource budgets. PMF-CL fits each task-local model to convergence, while the baselines train for one epoch. On Sequential MNIST, it also uses about 16 times as much memory as ER. This makes it hard to tell how much of the improvement comes from the method itself.

  2. The experiments are limited to synthetic linear regression and a linear classifier on Sequential MNIST. They do not show how performance changes with the number of tasks, model dimension, or degree of task conflict.

  3. Eq. (7) and Algorithm 3 omit the 1/nk1/n_k normalization factor in the new-task terms, whereas Eq. (15) includes it. With unequal task sizes, this changes the effective task weights and can prevent recovery of the stated joint-training solution.

Questions for authors
  1. How does PMF-CL compare with ER and DER when they receive the same total memory budget and are allowed to train to convergence?

  2. How do the authors ensure that the synthetic tasks have no common minimizer, and how does PMF-CL perform as the conflict between tasks increases?

Reviewer_RmNnOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper reformulates continual learning (CL) as a multi-task optimization problem with conflicting tasks. The authors propose PMF-CL, which preserves representations of past task losses to recover a specified hindsight joint-training solution. For quadratic losses, the method provides exact Pareto-optimality guarantees. For losses admitting quadratic upper bounds (QUBs), it achieves Pareto optimality for the quadratic surrogates and provides upper bounds on forgetting under the original losses.

Critical strengths

The paper provides a theoretically motivated perspective on catastrophic forgetting through Pareto optimization. It connects the trade-offs between conflicting tasks to the information needed to preserve past loss functions.

Critical weaknesses

The exact results focus on linear and basis-function regression, while the broader extension relies on quadratic upper bounds. The experiments are limited to synthetic linear regression and linear classification on Sequential MNIST. Although the introduction motivates CL for large-scale neural networks, the paper does not demonstrate the effectiveness of PMF-CL on such models.

The novelty relative to established incremental optimization methods remains unclear. Section 4.1 explicitly acknowledges that, under sample-proportional task weights, the linear regression update reduces to recursive least squares (RLS). The paper identifies the Pareto formulation and loss-preserving minimal sufficient information (MSI) as its contributions. However, it remains unclear what additional theoretical or algorithmic insight these provide beyond sufficient-statistic accumulation and the standard Pareto-optimality guarantee of globally minimizing a positively weighted sum of task losses. The relationship to curvature-based CL methods also needs clearer discussion.

Although the memory requirement is independent of the number of tasks, the general O(d2)O(d^2) storage cost can be prohibitive for large neural networks. The paper discusses Fisher-based approximations, but does not establish whether these approximations preserve the upper-bound property or the resulting forgetting guarantees. Its practical scalability therefore remains uncertain.

Questions for authors
  1. What additional insight does the MSI formulation provide beyond existing sufficient statistics and curvature-based representations? In particular, what is new about its minimality characterization?
  2. Given the stated equivalence to RLS in the linear regression setting, could the authors clarify which results go beyond this equivalence and provide theoretical or empirical comparisons with Online EWC and related incremental learning methods?
  3. How can PMF-CL scale to large neural networks given its O(d2)O(d^2) memory requirement? Under what conditions would practical curvature approximations preserve its theoretical guarantees?
  4. What practical advantages does PMF-CL offer over existing incremental optimization methods beyond linear regression?
Reviewer_2MEzOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper concentrates on what information needs to be stored to minimize catastrophic forgetting remains elusive in CL. I argues that in practice, tasks rarely have a common global minimizer, and a certain amount of forgetting is inevitable. This paper proposes a foundational reframing of CL as balancing conflicting tasks in hindsight from a multi-task learning (MTL) perspective, which is based on finding Pareto-optimal solutions. The main contributions can be summarized as:

  1. Propose a general meta-algorithmic framework for a Pareto-minimal-forgetting continual learner (PMF-CL) that can handle conflicting tasks with differing minimizers.

  2. Link CL and MTL for general loss functions and reframe CL as finding an MTL solution sequentially.

  3. Investigate the performances of PMF-CL in the case of quadratic and Quadratic Upper Bounded (QUB) loss functions.

  4. Characterize the memory required to achieve provably minimal forgetting.

  5. Perform numerical experiments for continual linear and logistic regression.

Critical strengths

1.This paper is well written and easy to follow.

2.Interesting ideas. This paper derives a PMF-CL algorithm for linear and basis-function regression and naturally extends to loss functions with quadratic upper-bounds around their minimizers.

  1. In Appendix G, this paper provides experiments to prove that the PMF-CL algorithm outperforms prior empirical methods while occupying lesser memory.
Critical weaknesses

1.This paper has limited novelty of the proposed algorithm. The PMF-CL algorithm for linear regression is essentially equivalent to the classical RLS method, as acknowledged by the authors. Although the paper provides a Pareto-optimality interpretation, it remains unclear what fundamental algorithmic or theoretical advances are achieved beyond existing online optimization methods.

2.In Section G.1, the baseline algorithms compared by the author were all from five years ago, making it difficult to prove the effectiveness of the algorithms.

3.In Figure 4a, the author has marked that the past data memory for all three methods is 0.1MB. However, why is the bar graph of PMF-CL higher than those of ER and DER?

Questions for authors

Please address the concerns in Critical Weaknesses.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.