Research map / Label-efficient and robust learning / Recognition and long-tail
Self-Calibrating BCIs: Ranking and Recovery of Mental Targets Without Labels
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 CURSOR, a label-free self-calibrating BCI that recovers a mental face target from unlabeled EEG and latent-space stimuli by scoring hypothesis targets with a shuffled-vs-aligned regression error ratio, and releases an EEG dataset (N=29).
Problem
Recovering a mental target (a face image) held in mind from paired EEG and image data without any labels or pre-trained decoders, i.e., self-calibration in continuous latent domains.
Approach
Scores each hypothesis target by constructing a hypothesis-specific distance dataset from the latent L2 distances, training an EEG regressor, and taking the ratio of shuffled-to-aligned RMSE; the score peaks at the true target, breaking the dependency between decoder and similarity function without labels.
Key results
- Linear regression CURSOR scores correlate with target distance at mean R = -0.77 ± 0.04 SD across targets, with mean top rank 6.64 of 60.
- Top-ranked faces averaged distance 2.9 to the target, within the perceptual indistinguishability range reported by humans (H-ID identification rate near chance at d ≤ 1.6).
- Optimization in reduced space converged to faces at mean distance 0.93 from the target, versus 24.09 (S-LR) and 22.51 (Dummy) for controls; label reconstruction RMSE 0.18 ± 0.08.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CURSOR scores to rank candidate faces with a Linear Regressor | CURSOR · Face hypothesis ranking | score-distance Pearson correlation: -0.77 ± 0.04 SD | Helped |
| Use Linear Regression rather than SVR or MLP as the estimator (ablation) | CURSOR · Face hypothesis ranking | ranking performance | Helped |
| Replace the original EEG representation with EEGNet embeddings (ablation) | CURSOR · Face hypothesis ranking | ranking performance | No clear effect |
| Optimize candidate faces using CURSOR (LR) | CURSOR · Reduced-space target optimization | optimized face distance to target: 24.09 -> 0.93; 22.51 -> 0.93 | Helped |
| Remove near-target stimulus-response pairs (ablation) | CURSOR (LR) · Reduced-space target optimization | optimized face distance to target | Mixed |
| Recover distance labels without supervision using CURSOR | CURSOR · EEG distance-label reconstruction | RMSE: 0.18 ± 0.08 vs 0.17 ± 0.09 | No clear effect |
Surprising findings
- A modern EEGNet embedding followed by an MLP performed equivalently to plain linear regression for the ranking task.
Limitations
- Restricted to synthetic faces from a single pre-trained GAN; generalization to real faces untested.
- Acquisition never explored latent distances beyond 40 and used log-spaced sampling, limiting closed-loop extension; evaluation is offline, not real-time.
- Optimization used a heuristically chosen 10-dimensional latent subspace, and cross-subject generalization remains untested.
Experimental setup
Datasets: CelebA-HQ Collected EEG-face RSVP dataset (N=9234 pairs, 29 participants)
Benchmarks: Ranking of 60 hypothesis faces against unknown target Optimization in PCA-reduced latent space Human identification study H-ID Human ranking study H-Rank
Metrics: Pearson correlation between score and distance target rank distance of top-ranked hypothesis RMSE of recovered labels perceptual identification rate
Baselines: Dummy estimator Shuffled linear regression (S-LR) Random guessing
Builds on: CMA-ES Optuna EEGNet progressive GAN
Closest papers by meaning
- NSD-Imagery: A Benchmark Dataset for Extending fMRI Vision Decoding Methods to Mental Imagery (CVPR 2025 · similarity 0.86)
- Bridging the Vision-Brain Gap with an Uncertainty-Aware Blur Prior (CVPR 2025 · similarity 0.85)
- MindAlign: Bridging EEG, Vision, and Language for Zero-Shot Visual Decoding (NeurIPS 2026 · similarity 0.85)
- MindEye2: Shared-Subject Models Enable fMRI-To-Image With 1 Hour of Data (ICML 2024 · similarity 0.85)
- Brain decoding: toward real-time reconstruction of visual perception (ICLR 2024 · similarity 0.85)
- $i$MIND: Insightful Multi-subject Invariant Neural Decoding (NeurIPS 2025 · similarity 0.85)
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.