Research map / Continual learning / Online and replay-based
Separating the 'what' and 'how' of compositional computation to enable reuse and continual learning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes a two-system "what/how" framework: an online incremental-EM task model infers time-varying epoch contexts from compositional cognitive tasks, which gate low-rank RNN components, enabling continual learning without catastrophic forgetting.
Problem
Continual learning and compositional reuse of computations in RNNs remain hard: sequentially trained RNNs suffer catastrophic forgetting, and existing methods do not explicitly reuse building blocks across compositional neuroscience tasks.
Approach
A probabilistic generative model of compositional cognitive tasks built from shared discrete task epochs; an online incremental EM learns this model trial-by-trial and infers epoch contexts that gate the weighted sum of low-rank recurrent components (U_z V_z^T) and input/output weights of an RNN.
Key results
- Context-modulated RNN mitigates catastrophic forgetting across sequentially trained tasks, maintaining high performance on all previous tasks, while Adam, EWC, neuromodulated RNN and hypernetwork baselines forget.
- Compositional generalization: the task model reaches ~83% accuracy on the new MemoryAnti task within ~40 trials, versus 56% (Adam), 53% (OWP), 64% (hypernetwork) after 512 trials of full training.
- Positive forward transfer between nearly all task pairs and positive backward transfer when tasks are trained for fewer trials.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use inferred computational contexts to gate and compose RNN components | ContextRNN · Sequential learning of six compositional cognitive tasks | test performance | Helped |
| Use inferred computational contexts to gate and compose RNN components | ContextRNN · MemoryAnti compositional generalization | accuracy: 83% within 40 trials; baselines: 56%, 53%, 64% after 512 trials | Helped |
| Train on a task with overlapping computational contexts before a new task | ContextRNN · Forward transfer across compositional cognitive tasks | test loss | Helped |
| Train on Delay or DM tasks before MemoryPro/Anti | ContextRNN · MemoryPro/Anti forward transfer | final test loss | Mixed |
| Continue training on a subsequent task that revisits shared contexts | ContextRNN · Backward transfer across compositional cognitive tasks | test loss on the previous task | Helped |
| Train subsequent tasks with OWP | OWP · Backward transfer across compositional cognitive tasks | test loss on old tasks | No clear effect |
| Train with Adam on a general RNN without continual-learning measures | General RNN · Sequential learning of six compositional cognitive tasks | test performance | Hurt |
| Use EWC on a general RNN | General RNN · Sequential learning of six compositional cognitive tasks | test performance | Hurt |
| Use OWP on a general RNN | General RNN · Sequential learning of six compositional cognitive tasks | task performance | Mixed |
| Use a jointly trained neuromodulatory subnetwork to control recurrent weights | Neuromodulated RNN · Sequential learning of six compositional cognitive tasks | test performance | Hurt |
| Use task embeddings and a hypernetwork to generate task-dependent weights | Hypernetwork RNN · Sequential learning of six compositional cognitive tasks | test performance | Mixed |
Surprising findings
- Standard EWC and the neuromodulated RNN baseline failed to prevent forgetting despite having a dedicated modulatory component.
- The unconstrained baseline RNNs have roughly twice as many trainable parameters as the context-modulated RNN, yet still underperform.
Limitations
- Context can only be inferred when it is observable from input-target pairs; epochs with identical observations (fixation and memory) are merged.
- Segmentation into epochs is not unique and depends on the modeling choices.
- Trial variable x is assumed discrete with a known number of values N_x.
Experimental setup
Datasets: synthetic compositional cognitive tasks (generated from the proposed generative model)
Benchmarks: synthetic cognitive task set (DelayPro/Anti, MemoryPro/Anti, DMPro/Anti, M'Pro/M'Anti)
Metrics: test log loss task performance accuracy forward transfer backward transfer trials to reach accuracy in compositional generalization
Baselines: Adam on general RNN EWC OWP neuromodulated RNN task-conditioned hypernetwork
Builds on: Adam Elastic Weight Consolidation Orthogonal Weight-Space Projection neuromodulated RNN task-conditioned hypernetwork
Closest papers by meaning
- A Second-Order Perspective on Model Compositionality and Incremental Learning (ICLR 2025 · similarity 0.87)
- Continual Learning with Global Alignment (NeurIPS 2024 · similarity 0.87)
- How does RL Post-training Induce Skill Composition? A Case Study on Countdown (NeurIPS 2026 · similarity 0.87)
- Scaling can lead to compositional generalization (NeurIPS 2025 · similarity 0.87)
- Attention as a Hypernetwork (ICLR 2025 · similarity 0.87)
- A Unified and General Framework for Continual Learning (ICLR 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.