Research map / Architectures and efficiency / RNNs and normalization
Temporal Latent Bottleneck: Synthesis of Fast and Slow Processing Mechanisms in Sequence Learning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 Temporal Latent Bottleneck to ViT | ViT · CIFAR10 and CIFAR100, 64 × 64 and 128 × 128 | accuracy: CIFAR10 64 × 64: 93.75 -> 94.79; CIFAR10 128 × 128: 73.18 -> 84.38; CIFAR100 64 × 64: 69.53 -> 79.17; CIFAR100 128 × 128 | Mixed |
| Add Temporal Latent Bottleneck to Decision Transformer | Decision Transformer · Atari (Breakout, Pong, Qbert, Seaquest) | score: Breakout: 71.51 -> 87.63; Pong: 13.68 -> 14.71; Qbert: 3268 -> 5019.75; Seaquest: 1039.11 -> 1248.22 | Helped |
| Add Temporal Latent Bottleneck to Transformer | Transformer · ListOps | accuracy: 37.64 -> 38.2 | Helped |
| Add Temporal Latent Bottleneck to Transformer | Transformer · Text Classification | accuracy: 64.0 -> 82.08 | Helped |
| Remove access to past information and top-down communication (ablation) | ViT + TLB · CIFAR10 64 × 64 | accuracy: 94.79 -> 91.30 | Hurt |
| Remove access to past information and top-down communication (ablation) | ViT + TLB · CIFAR10 128 × 128 | accuracy: 84.38 -> 72.92 | Hurt |
| Remove top-down conditioning (ablation) | ViT + TLB · CIFAR10 64 × 64 | accuracy: 94.79 -> 93.75 | Hurt |
| Remove top-down conditioning (ablation) | ViT + TLB · CIFAR10 128 × 128 | accuracy: 84.38 -> 83.59 | Hurt |
| Replace TLB memory with direct attention to previous few chunks (ablation) | Transformer + TLB · ListOps | accuracy: 38.2 -> 32.10 | Hurt |
| Remove top-down communication by using layer-specific TLBs (ablation) | Transformer + TLB · ListOps | accuracy: 38.2 -> 37.57 | Hurt |
| Vary chunk size (ablation) | Transformer + TLB · ListOps | accuracy: Chunk size 2: 36.08; 20: 36.49; 100: 38.18; 500: 36.97; 1000: 36.82 | Mixed |
| Vary the number of temporal latent bottleneck state vectors (ablation) | Transformer + TLB · ListOps | accuracy: 1: 36.42; 10: 37.16; 20: 38.18; 200: 37.25 | Mixed |
Closest papers by meaning
- Understanding and Controlling Memory in Recurrent Neural Networks (ICML 2019 · similarity 0.88)
- Before It Fades: Reinforcing Temporal Representations at Inference Time in VideoLLMs (NeurIPS 2026 · similarity 0.86)
- Auto-Compressing Networks (NeurIPS 2025 · similarity 0.85)
- Forgetting Transformer: Softmax Attention with a Forget Gate (ICLR 2025 · similarity 0.85)
- When Do Transformers Outperform Feedforward and Recurrent Networks? A Statistical Perspective (NeurIPS 2025 · similarity 0.85)
- Algorithm Development in Neural Networks: Insights from the Streaming Parity Task (ICML 2025 · similarity 0.85)
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.