My Blog

I started by converting raw EEG channels into normalized temporal windows, then applied a lightweight CNN to learn local spatial patterns across channels. The BiLSTM block followed to capture forward and backward temporal context, which improved class separation for subtle mental state transitions.

The most impactful change was balancing precision and recall using class-weighted loss, because the "focus" class had fewer examples. In cross-subject testing, the hybrid model outperformed standalone CNN and LSTM baselines and gave more stable results across sessions.

Our team built a pipeline that scanned incoming datasets, compared target distribution across demographic groups, and flagged parity gaps using statistical thresholds. We then generated a plain-language bias report with suggestions like resampling, feature review, and threshold calibration.

The biggest lesson from the event was that fairness tooling must be easy enough for non-ML stakeholders to use. By converting metrics into actionable recommendations, we made the output useful for both engineers and decision makers.