Videos
Short explainers on my research.
2026
- Can clinical context improve recognition of impaired speech? Across ten frozen audio-language models the answer is mostly no — added context often makes word error rate worse. This explains the counterfactual controls behind that result, why training exposure removes the penalty, and where context still helps: the hardest utterances.
- The full explainer. Why calibrating a visual cortical prosthesis electrode by electrode does not scale, and how we instead train networks on trial-resolved recordings to shape stimulation-evoked population activity — reaching neural targets at lower stimulation currents, and producing more consistent percepts, than conventional methods.
- A condensed version of the same work for a general audience: what it takes to control neural activity in the human visual cortex through a bidirectional implant, and why that matters for restoring sight.
- Speech recognition fails the people who most need it, with error rates above 40% for impaired speech. This explains how we adapt a foundation model to one speaker using Bayesian low-rank adaptation, cutting character error rate roughly in half while training far fewer parameters.