The readout-decomposition of activation change (RDAC) framework analyzes the stability-plasticity dilemma in continual learning through the geometry of prior-task readouts: learning-induced activation change decomposes into a component in the range of a prior readout, which can alter prior-task outputs and hence threatens stability, and a component in the orthogonal null space, which cannot affect them; only the former is associated with forgetting, while restricting the latter costs plasticity. The framework's exact instrument, a gradient decomposition algorithm that confines hidden-weight updates to the null space of prior readouts, is derived only for one-hidden-layer linear networks under task-incremental learning with frozen readouts. The open problem, raised in separate future-work remarks by the source authors, is to extend this exact analysis and its algorithmic use to deep and nonlinear networks (the source frames this as a search for solutions, possibly approximate, that might guide a generic gradient-decomposition-based algorithm, and notes that an analytical treatment of non-linear settings would shed light on the learning dynamics of replay-based methods), to continual-learning scenarios beyond frozen per-task readouts, and to test the hypothesis that subspace-level (Hopfieldian) regularization preserves more plasticity at matched stability than per-weight (Sherringtonian) importance scores such as EWC and SI. The authors' own 2024 revision, published at CoLLAs 2024, resolved only the data-dependent diagnostic estimation of task-relevant subspaces in deep nonlinear networks; the analytical extension, the non-task-incremental extension, and the Hopfieldian regularization test remain open.

Background

Continual learning (CL) asks a neural network to learn a sequence of tasks without full retraining; the central difficulty is catastrophic forgetting, the loss of performance on earlier tasks as new ones are learned, in tension with the need to keep adapting. This tension is the stability-plasticity dilemma. A standard taxonomy (van de Ven, Tuytelaars and Tolias, 2022) distinguishes task-incremental learning, in which each task has its own output layer, task identity is known at test time, and these readouts are frozen while shared hidden weights keep learning, from class-incremental and domain-incremental scenarios, where task identity must be inferred or the same outputs persist across tasks. A readout is the weight matrix mapping hidden-layer activations to a task's outputs; for a prior task with readout WRW_R, the activation change induced by later learning decomposes into a component in the range of WRW_R (the row space, visible to the readout and able to change its outputs) and a component in the orthogonal null space (unable to affect prior outputs).

Popular CL algorithm families include per-weight importance regularization, such as Elastic Weight Consolidation (EWC; Kirkpatrick et al., 2017) and Synaptic Intelligence (SI; Zenke, Poole and Ganguli, 2017), which add a quadratic penalty on individual weights estimated to be important for prior tasks; replay-based methods, such as Gradient Episodic Memory (GEM; Lopez-Paz and Ranzato, 2017), which retain a small episodic memory of prior data and constrain new-task gradients not to increase prior-task loss; and gradient or null-space projection methods (Orthogonal Gradient Descent, Farajtabar, Azizan, Mott and Li, 2019; Gradient Projection Memory, Saha, Garg and Roy, 2021; advanced null-space methods, Kong, Liu, Wang and Tao, 2022), which estimate from old-task data the subspace of weight space that affects prior tasks and project new updates out of it.

The RDAC framework (Readout-Decomposition of Activation Change; Anthes, Thorat, König and Kietzmann, arXiv:2310.04741) uses the readout geometry above as a lens on all of these. Its findings: only activation change in the range of prior readouts is associated with forgetting, and per-weight regularization methods such as EWC and SI unnecessarily restrict change in the null space as well, paying a plasticity cost with no stability benefit, because per-weight importance scores are not aligned with the readout geometry; replay-based methods largely preserve both. For one-hidden-layer linear networks, the framework yields an exact gradient decomposition algorithm: by scaling and projecting the backpropagated error signal (parameters α\alpha and β\beta control how much hidden-weight update is allowed in the range and null space of the prior readout), learning can be confined to the null space of prior readouts, maintaining stability at minimal plasticity cost. The framework is also proposed as a lens on representational drift in biological neural circuits, where neural representations change over time while behavior is preserved.

Two further terms from the source's discussion: Barack and Krakauer (2021) distinguish a Sherringtonian view of neural computation, which describes processing at the level of individual synapses (weights), from a Hopfieldian view, which describes it at the level of population activity and the representation subspaces it spans. EWC and SI are Sherringtonian (importance per weight); the source hypothesizes that a Hopfieldian regularization scheme, assigning importance to subspaces — concretely, in the source's proposal, 'a regularization algorithm relying on the principal dimensions of the learning in weight space' — could preserve more plasticity at the same stability, but this has not been tested.

Everything exact in RDAC is confined to one-hidden-layer linear networks and to task-incremental learning with frozen readouts. With multiple hidden layers the allowed update of each layer depends on the updates of all upstream layers, so the exact null-space constraints couple across layers (the source demonstrates this complication even for deep linear networks); with nonlinear activations, a projection computed on the readout no longer acts directly on hidden-weight updates. In its 2024 revision (published at the 3rd Conference on Lifelong Learning Agents, CoLLAs 2024, under the title 'Keep Moving: identifying task-relevant subspaces to maximise plasticity for newly learned tasks'), the authors added a data-dependent approximation, estimating per layer a 'functional range' and 'functional nullspace' by singular value decomposition of sampled old-task loss gradients, and used it to show that the linear-case findings generalize to a deep nonlinear network on Split CIFAR-10 (a continual-learning benchmark that splits the CIFAR-10 image-classification dataset into sequential tasks); this approximation requires access to prior-task data, is valid only locally in weight space, and leaves the analytical questions open.

Problem Statement

Extend the RDAC framework's exact, readout-geometry-based account of the stability-plasticity trade-off beyond the one-hidden-layer linear, task-incremental setting in which it was derived. The common objective is to make the account general enough to analyze and design continual-learning algorithms for realistic networks. The source authors raise three separate open questions toward this objective, in different places of their discussion; each is an independent direction of research, not a conjunct of deliverables that a single answer must supply together: (1) Analytical extension of the gradient decomposition: the source states that the introduction of non-linearities and multiple hidden layers makes its exact procedure 'non-trivially complex and beyond the purvey of this paper', and leaves 'the search for approximate solutions that might guide a generic gradient-decomposition-based algorithm as a quest for the future'. Within this question, an answer would derive gradient/weight-update decompositions that leave prior-task mappings unchanged in networks with multiple hidden layers and non-linear activations, under the source's baseline of task-incremental learning with frozen per-task readouts (the source shows exact null-space constraints couple across layers even in deep linear networks), where 'unchanged' carries the source's sense that an activation change with no component in the range of a frozen prior readout cannot alter that task's outputs. Consistent with the source's own framing, principled approximate or locally valid decompositions that might guide a generic gradient-decomposition-based algorithm fall within this question, as do rigorous characterizations of what such decompositions can and cannot achieve. The source motivates this extension in part by noting that an analytical treatment of non-linear settings 'would be crucial to shed light on the learning dynamics' of replay-based continual-learning algorithms, which the one-hidden-layer linear analysis cannot treat. (2) Extension beyond task-incremental learning with frozen prior readouts: adapt the framework to scenarios in which the readout layer remains plastic or is shared across tasks, possibly with context signals, and to the class-incremental and domain-incremental scenarios, where stability can no longer be defined by freezing prior readouts (the source notes it is 'not immediately clear how the RDAC framework could be extended to these scenarios'). (3) Hopfieldian versus Sherringtonian regularization: construct and empirically test a regularization scheme that assigns importance at the level of subspaces rather than individual weights (as EWC and SI do), and determine whether it preserves more plasticity at matched stability. Boundary against prior progress: the authors' CoLLAs 2024 revision already provides a data-dependent approximation of task-relevant subspaces in deep non-linear networks (functional range/nullspace from sampled old-task gradients); this approximation is prior progress toward question (1) rather than an answer to it, and re-deriving it does not resolve any of the open questions above.

References

  1. Daniel Anthes, Sushrut Thorat, Peter König, Tim C. Kietzmann. Balancing stability and plasticity in continual learning: the readout-decomposition of activation change (RDAC) framework (v1, 2023); revised as Keep Moving: identifying task-relevant subspaces to maximise plasticity for newly learned tasks (v6, 2024, published at the 3rd Conference on Lifelong Learning Agents, CoLLAs 2024). DOI: 10.48550/arXiv.2310.04741. URL: https://arxiv.org/abs/2310.04741
  2. James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A. Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, Demis Hassabis, Claudia Clopath, Dharshan Kumaran, Raia Hadsell. Overcoming catastrophic forgetting in neural networks. Proceedings of the National Academy of Sciences 114(13):3521-3526, 2017. DOI: 10.1073/pnas.1611835114. URL: https://doi.org/10.1073/pnas.1611835114
  3. Friedemann Zenke, Ben Poole, Surya Ganguli. Continual Learning Through Synaptic Intelligence. 2017. arXiv:1703.04200. URL: https://arxiv.org/abs/1703.04200
  4. David Lopez-Paz, Marc'Aurelio Ranzato. Gradient Episodic Memory for Continual Learning. 2017. arXiv:1706.08840. URL: https://arxiv.org/abs/1706.08840
  5. Mehrdad Farajtabar, Navid Azizan, Alex Mott, Ang Li. Orthogonal Gradient Descent for Continual Learning. 2019. arXiv:1910.07104. URL: https://arxiv.org/abs/1910.07104
  6. Gobinda Saha, Isha Garg, Kaushik Roy. Gradient Projection Memory for Continual Learning. 2021. arXiv:2103.09762. URL: https://arxiv.org/abs/2103.09762
  7. Yajing Kong, Liu Liu, Zhen Wang, Dacheng Tao. Balancing Stability and Plasticity through Advanced Null Space in Continual Learning. 2022. arXiv:2207.12061. URL: https://arxiv.org/abs/2207.12061
  8. Gido M. van de Ven, Tinne Tuytelaars, Andreas S. Tolias. Three types of incremental learning. Nature Machine Intelligence 4(12):1185-1197, 2022. DOI: 10.1038/s42256-022-00568-3. URL: https://doi.org/10.1038/s42256-022-00568-3
  9. David L. Barack, John W. Krakauer. Two views on the cognitive brain. Nature Reviews Neuroscience 22:359-371, 2021. DOI: 10.1038/s41583-021-00448-6. URL: https://doi.org/10.1038/s41583-021-00448-6
  10. Harel Yadid, Meir Yossef Levi, Roy Betser, Guy Gilboa. Make it SING: Analyzing Semantic Invariants in Classifiers. 2026. arXiv:2603.14610. URL: https://arxiv.org/abs/2603.14610