Research map / Probabilistic and scientific modeling / Dynamical systems
Exploring the trade-off between deep-learning and explainable models for brain-machine interfaces
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Adapts KalmanNet (a Kalman filter with RNN-computed gain) as a BMI finger decoder and compares it offline and online against a Kalman filter, tcFNN and LSTM in two monkeys, plus a heteroscedastic KF that mimics its trust behavior.
Problem
Deep-learning BMI decoders achieve state-of-the-art performance but are black boxes, raising safety concerns for physical control, while explainable Kalman filters underperform.
Approach
Use KalmanNet, which keeps the KF structure but replaces the Kalman gain computation with GRUs, yielding a input-dependent, non-linear trust between neural observations and dynamics; the authors then show a heteroscedastic KF with velocity-dependent process noise reproduces its performance while remaining explainable.
Key results
- Offline, KalmanNet matched the LSTM on velocity correlation (not significant, p = 0.64) and significantly outperformed the KF in velocity MSE (p < 0.01).
- Online over five days, KalmanNet achieved 97% success rate versus 90% for LSTM, with smoother paths, but lower throughput than LSTM (p < 1e-16).
- A heteroscedastic KF matched KalmanNet velocity performance (correlation 0.64 vs 0.65; MSE 3.4e-7 vs 3.3e-7), but assumes ground-truth velocity so is not usable online.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use KalmanNet instead of LSTM | Offline finger movement prediction · Offline finger movement prediction | velocity correlation | No clear effect |
| Use KalmanNet instead of LSTM | Offline finger movement prediction · Offline finger movement prediction | velocity MSE | No clear effect |
| Use KalmanNet instead of the Kalman filter | Offline finger movement prediction · Offline finger movement prediction | velocity correlation | Helped |
| Use KalmanNet instead of the Kalman filter | Offline finger movement prediction · Offline finger movement prediction | velocity MSE | Helped |
| Use KalmanNet instead of the Kalman filter or tcFNN | Online finger control · Online finger control | success rate: 97% vs. 90% | Helped |
| Use KalmanNet instead of LSTM | Online finger control · Online finger control | throughput | Hurt |
| Use KalmanNet instead of LSTM | Online finger control · Online finger control | trial time | Hurt |
| Use KalmanNet instead of LSTM | Online finger control · Online finger control | path efficiency | Helped |
| Add process noise that varies with ground-truth velocity | Kalman filter (HKF) · Offline finger movement prediction | velocity correlation | Helped |
| Add process noise that varies with ground-truth velocity | Kalman filter (HKF) · Offline finger movement prediction | velocity MSE: 3.4E −7 vs. 3.3E −7 | No clear effect |
| Use KalmanNet instead of LSTM | Offline decoding under injected noise · Noise magnitude 100 times training-data standard deviation | velocity MSE increase: ∼4300 vs. ∼2× | Hurt |
| Train on the normal task context rather than the spring+wrist context | KalmanNet · Unseen spring+wrist task context | velocity MSE increase: 44.45% | Hurt |
Surprising findings
- KalmanNet had the largest MSE increase under out-of-distribution noise (about 4300x at 100x noise magnitude), while the LSTM changed little (about 2x).
- Under task-context change, the KF showed the smallest MSE increase (8.0%) versus 34-50% for neural network decoders.
Limitations
- KalmanNet inherits the limited generalization of deep decoders across task contexts.
- Performance degrades sharply under unseen out-of-distribution noise because the KF inductive bias is retained.
- Only two monkeys and online experiments limited to Monkey N with a small number of days; tcFNN evaluated without ReFIT.
Experimental setup
Datasets: in-house monkey microelectrode recordings (SBP) during 2-DoF finger task
Benchmarks: in-house offline finger kinematics decoding (13 days) in-house online finger BMI task (5 days)
Metrics: Pearson correlation MSE success rate throughput trial time path efficiency normalized velocity MSE under injected noise off-context vs on-context MSE increase
Baselines: Kalman filter tcFNN LSTM heteroscedastic Kalman filter
Builds on: KalmanNet Kalman filter tcFNN LSTM
Closest papers by meaning
- Balancing memorization and generalization in RNNs for high performance brain-machine Interfaces (NeurIPS 2023 · similarity 0.90)
- A Scalable, Causal, and Energy Efficient Framework for Neural Decoding with Spiking Neural Networks (NeurIPS 2025 · similarity 0.85)
- SPINT: Spatial Permutation-Invariant Neural Transformer for Consistent Intracortical Motor Decoding (NeurIPS 2025 · similarity 0.85)
- Few-shot Algorithms for Consistent Neural Decoding (FALCON) Benchmark (NeurIPS 2024 · similarity 0.84)
- Neural-Behavioral Representation of Natural Whole-body Movement in Monkeys (NeurIPS 2026 · similarity 0.84)
- NeurIPT: Foundation Model for Neural Interfaces (NeurIPS 2025 · similarity 0.83)
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.