Research map / Architectures and efficiency / Attention and transformers
Learning to Combine Top-Down and Bottom-Up Signals in Recurrent Neural Networks with Attention over Modules
Sarthak Mittal, Alex Lamb, Anirudh Goyal, Vikram Voleti, Murray Shanahan, Guillaume Lajoie, Michael Mozer, 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 |
|---|---|---|---|
| Add bidirectional top-down attention to the hierarchical modular recurrent architecture (ablation) | Hierarchical RIMs · Sequential MNIST 19×19 | test accuracy (%): 72.0 -> 74.2 | Improved |
| Add bidirectional top-down attention to the hierarchical modular recurrent architecture (ablation) | Hierarchical RIMs · Sequential MNIST 24×24 | test accuracy (%): 50.3 -> 51.4 | Improved |
| Add bidirectional top-down attention to the hierarchical modular recurrent architecture (ablation) | Hierarchical RIMs · Sequential CIFAR 32×32 | test accuracy (%): 46.8 -> 52.2 | Improved |
| Add bidirectional top-down attention to the hierarchical modular recurrent architecture (ablation) | Hierarchical RIMs · WikiText-103 test | perplexity: 38.1 -> 36.8 | Improved |
| Add bidirectional top-down attention to the hierarchical modular recurrent architecture (ablation) | Hierarchical RIMs · Sequential CIFAR 19×19 | test accuracy (%): 57.2 -> 60.1 | Improved |
| Use bidirectional top-down information flow (ablation) | LSTM with hierarchy · Sequential CIFAR 32×32 | test accuracy (%): 33.2 -> 42.1 | Improved |
| Add attention to the hierarchical LSTM (ablation) | LSTM with hierarchy · Sequential CIFAR 32×32 | test accuracy (%): 33.2 -> 40.0 | Improved |
| Add bidirectional top-down connections to the hierarchical attention LSTM (ablation) | LSTM with hierarchy and attention · Sequential CIFAR 32×32 | test accuracy (%): 40.0 -> 43.0 | Improved |
| Add bidirectional top-down connections to the hierarchical attention LSTM (ablation) | LSTM with hierarchy and attention · Sequential MNIST 24×24 | test accuracy (%): 31.0 -> 20.8 | Hurt |
| Use BRIMs instead of LSTM | Recurrent policy network · Atari Asteroids | score: 3801 -> 2000 | Hurt |
| Use BRIMs instead of LSTM | Recurrent policy network · Atari Alien | score: 1612 -> 4102 | Improved |
| Use BRIMs instead of LSTM | Recurrent model · Adding, 1000 train/2000 test sequence length, 10 values | error: 17.246 -> 11.90 | Improved |
Closest papers by meaning
- Understanding the Differences in Foundation Models: Attention, State Space Models, and Recurrent Neural Networks (NeurIPS 2024 · similarity 0.85)
- Mind the Gap: a Spectral Analysis of Rank Collapse and Signal Propagation in Attention Layers (ICML 2025 · similarity 0.85)
- Hierarchical Self-Attention: Generalizing Neural Attention Mechanics to Multi-Scale Problems (NeurIPS 2025 · similarity 0.84)
- Self-Attention-Based Contextual Modulation Improves Neural System Identification (ICLR 2025 · similarity 0.84)
- Contrastive Consolidation of Top-Down Modulations Achieves Sparsely Supervised Continual Learning (NeurIPS 2025 · similarity 0.84)
- In-Context Denoising with One-Layer Transformers: Connections between Attention and Associative Memory Retrieval (ICML 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.
