Research map / Architectures and efficiency / RNNs and normalization
A Modern Self-Referential Weight Matrix That Learns to Modify Itself
Kazuki Irie, Imanol Schlag, Róbert Csordás, Jürgen Schmidhuber
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use self-modifying SRWM instead of DeltaNet | few-shot classifier · Mini-ImageNet 5-way 5-shot | accuracy: 62.7 ± 0.1 -> 61.4 ± 0.1 | Hurt |
| use self-modifying SRWM instead of DeltaNet | sequential adaptation model · Mini-ImageNet after Omniglot | total accuracy: 50.4 -> 53.3 | Improved |
| use self-modifying SRWM instead of DeltaNet | ProcGen agent · 6 environments easy, test | normalized aggregated score: 18.6 (1.7) -> 20.0 (1.8) | Improved |
| remove self-modification (ablation) | SRWM ProcGen agent · 6 environments easy, test | normalized aggregated score: 20.0 (1.8) -> 15.3 (1.9) | Hurt |
| add persistent weight updates across the episode instead of resetting every backpropagation span (ablation) | SRWM ProcGen agent · 6 environments easy, test | normalized aggregated score: 16.1 (2.2) -> 20.0 (1.8) | Improved |
| replace DeltaNet slow weights with SRWM | DeltaNet · ProcGen 4 environments memory, train | normalized aggregated score: 51.8 (2.6) -> 59.0 (2.1) | Improved |
| replace DeltaNet slow weights with SRWM | DeltaNet · ProcGen 4 environments memory, test | normalized aggregated score: 38.0 (4.1) -> 38.5 (3.2) | No clear effect |
| use SRWM instead of DeltaNet | few-shot classifier · Omniglot 5-way 1-shot | accuracy: 97.2 ± 0.0 -> 97.4 ± 0.0 | No clear effect |
| use SRWM instead of DeltaNet | few-shot classifier · Mini-ImageNet 5-way 1-shot | accuracy: 47.0 ± 0.1 -> 47.0 ± 0.2 | No clear effect |
| use SRWM instead of DeltaNet | few-shot classifier · FC100 5-way 5-shot | accuracy: 43.2 ± 0.1 -> 44.5 ± 0.1 | No clear effect |
| increase vision backend from Conv-4-32 to Res-12-256 (ablation) | SRWM · FC100 5-way 5-shot | accuracy: 44.5 ± 0.1 -> 46.3 ± 0.1 | Improved |
| use Res-12-512 instead of Res-12-256 (ablation) | SRWM · FC100 5-way 5-shot | accuracy | No clear effect |
Closest papers by meaning
- Linear Transformers Are Secretly Fast Weight Programmers (ICML 2021 · similarity 0.85)
- Weight Agnostic Neural Networks (NeurIPS 2019 · similarity 0.85)
- 1000 Layer Networks for Self-Supervised RL: Scaling Depth Can Enable New Goal-Reaching Capabilities (NeurIPS 2025 · similarity 0.84)
- RECAST: Reparameterized, Compact weight Adaptation for Sequential Tasks (ICLR 2025 · similarity 0.84)
- Training-Free Task Vectors for LLM Behavioral Control (NeurIPS 2026 · similarity 0.84)
- Self-Training with Dynamic Weighting for Robust Gradual Domain Adaptation (NeurIPS 2025 · similarity 0.84)
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.
