Research map / Label-efficient and robust learning / Noisy labels
DeCorrNet: Enhancing Neural Decoding Performance by Eliminating Correlations in Noise
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 DeCorrNet, an end-to-end denoising module that learns to predict and subtract correlated noise from neural recordings using a decorrelation loss, improving downstream neural decoders on BCI datasets.
Problem
Neural decoders for brain-computer interfaces typically assume independent noise, but correlated noise across neural channels degrades decoding performance.
Approach
Use a task encoder (external attention) and self-attention to predict noise from neural signals, subtract it, and train with a learnable-weighted decorrelation loss that minimizes noise correlations within each task while preserving signal correlations across tasks; the module is plug-and-play for any decoder.
Key results
- On the handwriting dataset, decorrelated data with iTransformer reached 96.1% accuracy versus 93.4% raw (+2.7%).
- On the speech dataset, decorrelated data improved GRU accuracy by 10.7 points (58.8% to 69.5%).
- DeCorrNet denoising outperformed six denoising baselines (e.g., Center-Out (M) SVM 85.5% vs. next best 75.6% FECAM).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| decorrelate neural data before decoding | DeCorrNet with iTransformer · Handwriting I | accuracy: 93.4 -> 96.1 | Helped |
| decorrelate neural data before decoding | DeCorrNet with GRU · Speech | accuracy: 58.8 -> 69.5 | Helped |
| decorrelate neural data before decoding | DeCorrNet with MLP · Center-Out (M) V | accuracy: 84.6 -> 95.1 | Helped |
| decorrelate neural data before decoding | DeCorrNet with iTransformer · Center-Out (RS) | accuracy: 95.8 -> 99.0 | Helped |
| add a Task Encoder rather than a fully connected network (ablation) | DeCorrNet · Handwriting I | accuracy: 93.6% -> 96.1% | Helped |
| add Self-Attention (ablation) | DeCorrNet · Handwriting I | accuracy: 94.5% -> 96.1% | Helped |
| use learned loss weights instead of fixed weights (ablation) | DeCorrNet · Handwriting I | accuracy: 93.1% -> 96.1% | Helped |
Surprising findings
- Some existing denoising methods (e.g., DeepFilter on Center-Out (M), wavelet transform on speech) degraded decoding performance below raw data.
Limitations
- Only eliminates noise correlations between neurons; correlations across time periods are not addressed.
- Uses only 20% of data for training due to limited clinical samples, with unclear generalization beyond the evaluated datasets.
Experimental setup
Datasets: simulated LNP neural dataset vision animals dataset (Kafashan et al. 2021) handwriting dataset (Willett et al. 2021) speech dataset (Willett et al. 2023) Center-Out (M) (Ma et al. 2023) Center-Out (RS) (Sundiang et al. 2023)
Benchmarks: simulated dataset vision animals dataset handwriting dataset speech dataset Center-Out (M) Center-Out (RS)
Metrics: classification accuracy discrimination score neuronal correlation coefficient
Baselines: Wavelet Transform Diffusion FECAM DeepFilter TCDAE MECG-E SVM GRU MLP ModernTCN TimesNet iTransformer
Builds on: external attention self-attention (Transformer) shuffling method for noise correlations
Closest papers by meaning
- SPINT: Spatial Permutation-Invariant Neural Transformer for Consistent Intracortical Motor Decoding (NeurIPS 2025 · similarity 0.85)
- A Scalable, Causal, and Energy Efficient Framework for Neural Decoding with Spiking Neural Networks (NeurIPS 2025 · similarity 0.85)
- Few-shot Algorithms for Consistent Neural Decoding (FALCON) Benchmark (NeurIPS 2024 · similarity 0.85)
- Leveraging unlabelled data for generalizable neural population decoding (NeurIPS 2026 · similarity 0.83)
- Generalizable, real-time neural decoding with hybrid state-space models (NeurIPS 2025 · similarity 0.83)
- NeurIPT: Foundation Model for Neural Interfaces (NeurIPS 2025 · similarity 0.82)
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.