Research map / Architectures and efficiency / RNNs and normalization
Scalable Real-Time Recurrent Learning Using Columnar-Constructive Networks
Khurram Javed, Haseeb Shah, Richard S. Sutton, Martha White
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 |
|---|---|---|---|
| combine columnar and constructive networks into CCNs | RTRL LSTM learners · Trace patterning | prediction error | Improved |
| CCNs instead of T-BPTT | LSTM state construction · Atari prediction 16x16 | normalized return error: over 5x lower error in many environments; Breakout error 20% more | Improved |
| add online feature normalization to CCNs (ablation) | CCN · Trace patterning | prediction error | Improved |
| add online feature normalization to CCNs (ablation) | CCN · Atari prediction 16x16 (10M steps) | prediction error | Improved |
| increase T-BPTT truncation length with fixed 8 features (ablation) | T-BPTT · Atari prediction 16x16 | prediction error: truncation 2 to 15 | Improved |
| increase T-BPTT feature count with truncation 8 (ablation) | T-BPTT · Atari prediction 16x16 | prediction error: 2 features error twice that of 15 features | Improved |
| large networks with short truncation (3:10, 5:8) vs small networks with long truncation (15:4) (ablation) | T-BPTT · Trace patterning | prediction error: 3:10 and 5:8 perform worst; 15:4, 30:2, 20:3 better | Hurt |
| allow T-BPTT 10 features with longer truncation (more compute) (ablation) | T-BPTT · Trace patterning | prediction error: nearly matches CCN at around seven times more compute | Improved |
| increase features-per-stage from 1 to 6 (ablation) | CCN · Trace patterning | prediction error: 1 to 6 features | Improved |
| reduce steps-per-stage to 0.5 million (ablation) | CCN (2 features-per-stage) · Trace patterning | asymptotic error: 0.5 million steps-per-stage | Mixed |
| add hierarchical features (CCN) over one-feature-at-a-time Constructive networks (ablation) | Constructive networks · Trace patterning | prediction error | Improved |
Closest papers by meaning
- Exploring the Promise and Limits of Real-Time Recurrent Learning (ICLR 2024 · similarity 0.92)
- Real-Time Recurrent Learning using Trace Units in Reinforcement Learning (NeurIPS 2024 · similarity 0.90)
- Real-Time Recurrent Reinforcement Learning (AAAI 2025 · similarity 0.87)
- Efficient Recurrent Off-Policy RL Requires a Context-Encoder-Specific Learning Rate (NeurIPS 2024 · similarity 0.86)
- Handling Delay in Real-Time Reinforcement Learning (ICLR 2025 · similarity 0.86)
- SRL: Scaling Distributed Reinforcement Learning to Over Ten Thousand Cores (ICLR 2024 · similarity 0.86)
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.
