Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning
Few-shot Algorithms for Consistent Neural Decoding (FALCON) Benchmark
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.
Introduces FALCON, a benchmark suite of five intracortical/songbird neural datasets with held-in and held-out sessions and an EvalAI-based evaluation of few-shot, zero-shot, and test-time adaptive iBCI decoders, seeded with baselines.
Problem
Intracortical brain-computer interface decoders degrade under neural nonstationarity and need frequent recalibration, but methods for robust decoding are compared ad hoc without standardized datasets or metrics.
Approach
A standardized benchmark with held-in (full training) and held-out (minimal few-shot calibration) sessions across movement and communication iBCI tasks, evaluated via a platform requiring only submitted code that produces causal, open-loop predictions on withheld data.
Key results
- Static (zero-shot) Wiener Filter drops up to 0.28 R2 on M1 and RNN drops up to 1.94 R2 on M1 from held-in to held-out sessions.
- On M1-A held-out, the best few-shot baselines reach ~0.59 R2 (NDT2 Multi) versus 0.78 for its held-in score, and the oracle is 0.78 held-out.
- On H2, CORP test-time adaptation achieves 0.11 held-out WER versus 0.37 zero-shot and 0.03 for the oracle RNN Multi + LM.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use a multi-session NDT2 instead of a Wiener Filter with oracle data | movement decoding · M2 | R2: 0.27 vs 0.77 R2 WF/NDT2 Multi | Helped |
| add a language model prior | RNN Multi oracle decoder · H2 | WER: 0.11 vs 0.02 WER RNN Multi/+ LM | Helped |
| use supervised few-shot calibration rather than zero-shot decoding | movement decoders · M1-A | R2 | Helped |
| use test-time recalibration with pseudo-labels rather than zero-shot decoding | H2 RNN · H2 | WER: 0.53±0.02 -> 0.11±0.01 | Helped |
| transfer a static decoder to held-out sessions without recalibration | Wiener Filter · M1 | R2: drop ≤0.28 R2 | Hurt |
| transfer a static decoder to held-out sessions without recalibration | RNN · M1 | R2: drops up to 1.94 R2 | Hurt |
| apply few-shot unsupervised alignment | CycleGAN + WF and NoMAD + WF · M1-A | R2: at most a marginal improvement | Mixed |
| use zero-shot EnSongdec on held-out sessions rather than an oracle decoder | EnSongdec · B1 | MSE: 5.61 vs 21.8 | Hurt |
Surprising findings
- Few-shot baselines (NoMAD, CycleGAN) provide at most marginal improvement over the static Wiener Filter on M1, showing the difficulty of stabilizing decoders.
- Models trained on trialized data degraded significantly when evaluated continuously, indicating trial-structure memorization.
Limitations
- Open-loop prediction only; does not capture closed-loop iBCI control performance.
- Evaluation may reward models exploiting trial structure in the datasets.
- Limited subject diversity: multi-session data from few subjects, so subject generalization is not evaluated.
Experimental setup
Datasets: M1 M2 H1 H2 B1
Benchmarks: FALCON M1 FALCON M2 FALCON H1 FALCON H2 FALCON B1
Metrics: R2 word error rate mean squared error
Baselines: Wiener Filter RNN NDT2 Multi NoMAD + WF CycleGAN + WF RNN Multi + LM CORP EnSongdec
Builds on: NoMAD CycleGAN NDT2 CORP EnSongdec
Closest papers by meaning
- SPINT: Spatial Permutation-Invariant Neural Transformer for Consistent Intracortical Motor Decoding (NeurIPS 2025 · 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 0.87)
- Time-Masked Transformers with Lightweight Test-Time Adaptation for Neural Speech Decoding (NeurIPS 2025 · similarity 0.87)
- Leveraging unlabelled data for generalizable neural population decoding (NeurIPS 2026 · similarity 0.86)
- BaRISTA: Brain Scale Informed Spatiotemporal Representation of Human Intracranial Neural Activity (NeurIPS 2025 · 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.