Does Denoising Matter? An Interactive Two-Brain Decoder That Shows What Artifacts Cost

This demo offers a hands-on exploration of a novel tool for quantifying and visualizing the impact of different kinds of EEG artifacts on brain decoding performance. For demonstration purposes, the tool will be integrated with a multi-modal decoding system for inference on dyadic interaction organized around group or individualized participation frameworks. This system reads synchronized EEG and eye-gaze from both members of the dyad and labels their interaction as Solo, Cooperative, or Competitive. A Vision Transformer encodes the paired gaze heatmaps. A dual-stream Transformer encodes the paired 32-channel EEG with inter-brain synchrony tokens. An uncertainty-aware fuzzy gating layer reads each modality's prediction entropy and emits a per-sample fusion weight alpha score, which indicates how far the model currently trusts the brain over the body. On 28 dyads, fusion reaches 94.9% accuracy, against 77.8% for gaze alone and 69.3% for EEG alone.

What attendees do: Participants can manipulate parameters related to artifact contamination and mitigation. Turning a dial injects eye blinks, jaw EMG, or an electrode pop – a large single-channel step transient caused by momentary loss of electrode contact – into one of the dyadic partner's EEG. Each artifact type has its own intensity control, so visitors can add them one at a time or in combination. As contamination increases, the following synchronized metrics are visualized: the distortion of raw EEG traces; the increase in the EEG branch's prediction entropy, which reflects model uncertainty; and the fusion weight alpha score, which shifts away from EEG to prioritize gaze as the gating layer down-weights the contaminated modality. The decoder's accuracy drops in real time, and the cost of the artifact becomes a number the model reports about itself.

To compare solutions to artifact contamination, users can select from four generations of artifact reduction methods – namely, Independent Component Analysis (ICA) (Jung et al, 1997), Online Recursive Independent Component Analysis (ORICA) (Hsu et al, 2015), Independent Component U-net (IC-U-Net) (Chang et al, 2023), and Artifact Removal Transformer (ART) (Chuang, Chang, Huang, & Bessas, 2025). Users can watch the decoding model regain its diminished confidence in brain signal and compare the relative success of each method.

Relevance: Large pretrained biosignal models inherit whatever noise survives preprocessing, and that dependency is rarely measured against downstream accuracy. We make this relationship visible and let people explore it. The demo also puts dyadic, two-brain decoding in front of an audience that mostly encounters single-subject BCI, and shows a fusion mechanism whose per-sample decisions are open to inspection.