Research map / Reinforcement learning / Visual and representation RL
Neural optimal feedback control with local learning rules
Johannes Friedrich, Siavash Golkar, Shiva Farashahi, Alexander Genkin, Anirvan M. Sengupta, Dmitri Chklovskii
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 an internal state model to predict through delayed sensory feedback | Bio-OFC on double-integrator control · Double integrator, LDS1 and LDS2, across sensory delays | control cost | Improved |
| Learn the controller from delayed observations without an internal state model | Model-free policy gradient · Double integrator, LDS1 and LDS2, across sensory delays | control cost | Hurt |
| Adapt the Kalman gain online as observation noise changes | Bio-OFC · Double integrator, LDS1 with changing observation noise | prediction MSE | Improved |
| Learn the system matrices and filter jointly online | Bio-OFC · Double integrator, LDS1 and LDS2 | prediction MSE | Improved |
| Learn the controller while identifying the system in closed loop | Bio-OFC · Double integrator, LDS1 and LDS2 | control cost | Improved |
| Use local learning rules by replacing Cᵀ with L (ablation) | Bio-OFC adaptive filter · Double integrator, LDS1 with changing observation noise | prediction MSE | No clear effect |
| Over-represent the latent state as 3-dimensional rather than 2-dimensional (ablation) | Bio-OFC · Double integrator, LDS2 | prediction MSE | No clear effect |
| Keep controller weights fixed and adapt only the estimator (ablation) | Bio-OFC · Reaching movements, force-field adaptation | trajectory match to human behavior | No clear effect |
| Use signal-dependent motor noise instead of additive noise (ablation) | Bio-OFC · Reaching movements, force-field adaptation | trajectory match to human behavior | No clear effect |
| Initialize the alignment eigenvalue at or below zero (ablation) | Bio-OFC adaptive filter · Double integrator, LDS1 | asymptotic prediction MSE | Hurt |
| Replace the internal state estimator with direct policy-gradient control from delayed observations | Gym-Fly agent · Gym-Fly | target overshoot | Hurt |
Closest papers by meaning
- Stochastic Optimal Control and Estimation with Multiplicative and Internal Noise (NeurIPS 2024 · similarity 0.89)
- Feedback control guides credit assignment in recurrent neural networks (NeurIPS 2024 · similarity 0.88)
- Learning Options for Compositional Motor Control with Adapter Banks (NeurIPS 2026 · similarity 0.85)
- World Models as Reference Trajectories for Rapid Motor Adaptation (NeurIPS 2025 · similarity 0.84)
- Exploring the trade-off between deep-learning and explainable models for brain-machine interfaces (NeurIPS 2024 · similarity 0.84)
- Model Based Inference of Synaptic Plasticity Rules (NeurIPS 2024 · 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.
