Research map / Probabilistic and scientific modeling / Dynamical systems
SPINT: Spatial Permutation-Invariant Neural Transformer for Consistent Intracortical Motor Decoding
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.
Proposes SPINT, a permutation-invariant transformer for intracortical motor decoding that infers unit identities from few-shot unlabeled calibration data via context-dependent positional embeddings, achieving gradient-free cross-session generalization on FALCON M1, M2, and H1.
Problem
Cross-session nonstationarity in intracortical BCI decoding: changing neural unit composition and tuning breaks decoders across sessions, and existing alignment methods need test-time labels or parameter updates.
Approach
Replace fixed unit identities with context-dependent positional embeddings inferred by a shared permutation-invariant ID encoder (mean-pooled over unlabeled calibration trials), feed them into a cross-attention decoder, and regularize with dynamic channel dropout that samples a random dropout rate each iteration.
Key results
- SPINT achieves cross-session R^2 of 0.66 on M1, 0.26 on M2, and 0.29 on H1, outperforming ZS and FSU baselines (best baseline NoMAD+WF 0.49, 0.20, 0.13).
- SPINT surpasses Wiener Filter oracles on all three datasets and NDT2 Multi FSS on M1 (0.66 vs 0.59).
- SPINT reaches latency ratios of 0.13 (M1), 0.13 (M2), and 0.14 (H1), below real-time threshold of 1.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use SPINT with context-dependent positional embeddings instead of fixed positional embeddings (ablation) | SPINT · M1, M2, and H1 cross-session | R² | Helped |
| Use SPINT with context-dependent positional embeddings instead of no positional embeddings (ablation) | SPINT · M1, M2, and H1 cross-session | R² | Helped |
| Add dynamic channel dropout (ablation) | SPINT · M1, M2, and H1 cross-session | R² | Helped |
| Use SPINT rather than the strongest zero-shot or few-shot unsupervised baseline | Cross-session motor decoding · M1 | R²: 0.49 ± 0.03 -> 0.66 ± 0.07 | Helped |
| Use SPINT rather than the strongest zero-shot or few-shot unsupervised baseline | Cross-session motor decoding · M2 | R²: 0.22 ± 0.06 -> 0.26 ± 0.13 | Helped |
| Use SPINT rather than the strongest zero-shot or few-shot unsupervised baseline | Cross-session motor decoding · H1 | R²: 0.12 ± 0.06 -> 0.29 ± 0.15 | Helped |
| Reduce the available population to 20% of its original size | SPINT · M1 cross-session | R²: 0.52 | Hurt |
| Increase the number of held-in training days | SPINT · M1, M2, and H1 cross-session | R² | Helped |
Surprising findings
- On M1, SPINT with a single calibration trial achieves performance close to using all calibration trials.
- SPINT outperforms the Wiener Filter oracle and NDT2 Multi few-shot supervised model on M1 without test-time labels or parameter updates.
Limitations
- Relies on labels during training to learn the end-to-end unit identifier, which may be limited in real-world iBCI.
- Evaluated only in silico on FALCON datasets from a single subject per task; not validated in closed-loop in vivo use.
- Modest absolute R^2 on M2 and H1 (0.26 and 0.29) with large standard deviations.
Experimental setup
Datasets: FALCON M1 FALCON M2 FALCON H1
Benchmarks: FALCON M1 FALCON M2 FALCON H1
Metrics: R^2 inference latency
Baselines: Wiener Filter RNN (LSTM) CycleGAN + WF NoMAD + WF NDT2 Multi
Builds on: DeepSets NDT2 NoMAD CycleGAN
Closest papers by meaning
- Few-shot Algorithms for Consistent Neural Decoding (FALCON) Benchmark (NeurIPS 2024 · similarity 0.91)
- A Scalable, Causal, and Energy Efficient Framework for Neural Decoding with Spiking Neural Networks (NeurIPS 2025 · similarity 0.88)
- Generalizable, real-time neural decoding with hybrid state-space models (NeurIPS 2025 · similarity 0.88)
- Population Transformer: Learning Population-level Representations of Neural Activity (ICLR 2025 · similarity 0.88)
- Leveraging unlabelled data for generalizable neural population decoding (NeurIPS 2026 · similarity 0.88)
- CSBrain: A Cross-scale Spatiotemporal Brain Foundation Model for EEG Decoding (NeurIPS 2025 · similarity 0.87)
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.