BCIwiki (bciwiki.com) — Researchers at Ulster University have added an explicit decision-confidence readout neuron to an echo state network (ESN), letting the model report confidence continuously as a classification decision forms, and found that this confidence signal discriminated correct from erroneous decisions in EEG-based motor imagery classification, according to a preprint posted to bioRxiv on September 4, 2026. The work has not been peer reviewed.
ESNs are efficient, neuro-inspired computational frameworks well suited to time-series data, but their decision confidence is typically quantified in limited ways. The proposed readout neuron is trained from the outputs of the decision readout and tracks confidence throughout the decision process rather than only at the final label.
In a simulated decision task, confidence activity increased with stimulus strength, linking greater discriminability to higher confidence. On EEG motor imagery data, confidence activity increased with decision accuracy and separated correct from error decisions, particularly in higher-performing participants. The authors describe this as reflecting human-like metacognition.
The authors say the approach enables continual monitoring of decision confidence and could support more trustworthy ESN decisions, particularly in biomedical applications. The authors are E. Lemoine, B. Lenfesty, U. K. N. Mudavath, S. Bhattacharyya and K. Wong-Lin.