Research map / Architectures and efficiency / RNNs and normalization
Sparse Attentive Backtracking: Temporal Credit Assignment Through Reminding
Nan Rosemary Ke, Anirudh Goyal ALIAS PARTH GOYAL, Olexa Bilaniuk, Jonathan Binas, Michael Mozer, Chris Pal, Yoshua Bengio
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 sparse attentive backtracking instead of truncated BPTT | LSTM on copying sequences · Copying T=300 | accuracy (last 10 characters): 24.4 -> 99.9 | Improved |
| Use sparse attentive backtracking instead of full BPTT | LSTM on CIFAR10 pixel-by-pixel classification · CIFAR10 | accuracy: 58.3 -> 64.5 | Improved |
| Use sparse attentive backtracking instead of full self-attention | LSTM on permuted MNIST · Permuted MNIST | accuracy: 97.9 -> 94.2 | Hurt |
| Train on T=100 and transfer to longer sequences with SAB rather than full-BPTT LSTM | LSTM on copying memory · Copying transfer T=200 | accuracy (last 10 digits): 34% -> 95% | Improved |
| Train on T=100 and transfer to longer sequences with SAB rather than full-BPTT LSTM | LSTM on copying memory · Copying transfer T=5000 | accuracy (last 10 digits): 12% -> 41% | Improved |
| Add mental updates that propagate gradients locally around retrieved memories (ablation) | SAB on copying T=100, k_top=1 · Copying T=100 | accuracy: 49.0 -> 57.9 | Improved |
| Add mental updates that propagate gradients locally around retrieved memories (ablation) | SAB on copying T=100, k_top=5 · Copying T=100 | accuracy: 98.3 -> 100.0 | Improved |
| Attend to all past events instead of a sparse subset (ablation) | SAB on copying T=100 · Copying T=100 | accuracy: 100.0 -> 40.5 | Hurt |
| Use SAB rather than LSTM with full BPTT | Character-level language modeling · Text8 test | BPC: 1.42 -> 1.47; 1.42 -> 1.44 | Mixed |
Closest papers by meaning
- Titans: Learning to Memorize at Test Time (NeurIPS 2025 · similarity 0.86)
- Emergent mechanisms for long timescales depend on training curriculum and affect performance in memory tasks (ICLR 2024 · similarity 0.86)
- Do Your Best and Get Enough Rest for Continual Learning (CVPR 2025 · similarity 0.86)
- Understanding and Improving Length Generalization in Recurrent Models (ICML 2025 · similarity 0.85)
- Memorization Sinks: Isolating Memorization during LLM Training (ICML 2025 · similarity 0.85)
- How Diffusion Models Memorize (NeurIPS 2026 · 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.
