Research map / Multi-agent systems and games / Human-AI and goal-conditioned
The Sensory Neuron as a Transformer: Permutation-Invariant Neural Networks for Reinforcement Learning
Yujin Tang, David Ha
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 AttentionNeuron instead of a fixed-input FNN under shuffled observations | CartPoleSwingUpHarder · 5 observations, shuffled | score: 38 ± 120 -> 471 ± 426 | Improved |
| Accept extra duplicated input channels (ablation) | AttentionNeuron · CartPoleSwingUpHarder, 10 observations | score: 472 ± 426 -> 471 ± 425 | No clear effect |
| Add five noisy channels (ablation) | AttentionNeuron · CartPoleSwingUpHarder, 5 observations + 5 noise | score: 472 ± 426 -> 461 ± 410 | No clear effect |
| Use AttentionNeuron instead of an FNN for shuffled observations | PyBullet Ant · shuffled observations | score: 232 ± 112 -> 2576 ± 75 | Improved |
| Train the permutation-invariant student with behavior cloning rather than ES (ablation) | PyBullet Ant · shuffled observations | score: 2576 ± 75 -> 2034 ± 948 | Hurt |
| Increase the downstream network size for behavior cloning (ablation) | PyBullet Ant AttentionNeuron student · shuffled observations | score: 2034 ± 948 -> 2579 ± 457 | Improved |
| Use AttentionNeuron to construct a latent representation from shuffled screen patches | Atari Pong policy with CNN · shuffled screens | score: 21 | Improved |
| Train with randomly removed screen patches, then provide more patches at test time (ablation) | Atari Pong AttentionNeuron policy · Atari Pong | score | Improved |
| Use AttentionNeuron with AttentionAgent instead of AttentionAgent alone | CarRacing · KOF background | score: -81 ± 4 -> 646 ± 189 | Improved |
| Use AttentionNeuron with AttentionAgent instead of AttentionAgent alone | CarRacing · Mt. Fuji background | score: -57 ± 38 -> 503 ± 152 | Improved |
| Reshuffle observations every 25 steps rather than never reshuffling (ablation) | AttentionNeuron policies · CartPole | score: 472 ± 426 -> 107 ± 146 | Hurt |
| Reshuffle observations every 25 steps rather than never reshuffling (ablation) | AttentionNeuron policies · PyBullet Ant | score: 2576 ± 75 -> 2053 ± 225 | Hurt |
Closest papers by meaning
- System Neural Diversity: Measuring Behavioral Heterogeneity in Multi-Agent Learning (JMLR 2025 · similarity 0.88)
- Neurons as Detectors of Coherent Sets in Sensory Dynamics (NeurIPS 2025 · similarity 0.87)
- What should a neuron aim for? Designing local objective functions based on information theory (ICLR 2025 · similarity 0.86)
- Efficient Reinforcement Learning by Discovering Neural Pathways (NeurIPS 2024 · similarity 0.86)
- Deep Reinforcement Learning with Time-Scale Invariant Memory (AAAI 2025 · similarity 0.85)
- Adaptive Surrogate Gradients for Sequential Reinforcement Learning in Spiking Neural Networks (NeurIPS 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.
