2026 Theses Doctoral

Computational Tools and Data-Driven Analyses for Large-Scale Electrophysiology and Behavior

Yu, Han

Modern systems neuroscience routinely generates datasets — thousands of simultaneously recorded neurons, hours of high-resolution behavioral video — that exceed the capacity of the bespoke scripts and hand-curated pipelines on which the field has traditionally relied. The bottleneck on scientific discovery is increasingly not the recording itself but the analysis stack. This thesis develops computational tools for the two dominant data modalities of modern systems neuroscience — extracellular electrophysiology and behavioral video — and presents scientific work on how neural circuits vary across brain regions, how they mature during development, and how they fail in disease.

On the electrophysiology preprocessing side, DREDge is a decentralized motion-correction algorithm for high-density extracellular recordings that remains robust in regimes — deep insertions in non-human primates, chronic recordings with substantial population turnover — where the global-template approach underlying current spike sorters fails. Building on motion-corrected, spike-sorted units, NEMO learns a per-neuron representation by contrastively aligning extracellular waveforms with spike-time autocorrelograms; the resulting embedding outperforms supervised, variational, and feature-engineered baselines on visual-cortex cell-type classification (PV/SST/VIP) and on brain-region decoding across the IBL Brain-wide Map. On the behavior side, BEAST is a vision-transformer framework that learns frame-level embeddings of unlabeled neuroscience video through joint masked-reconstruction and temporal-contrastive pretraining, and transfers without per-task retraining to behavior segmentation, sample-efficient pose tracking, and neural encoding from facial and body video.

The thesis then turns to three scientific settings. An ongoing electrophysiological atlas of the mouse brain links large-scale extracellular electrophysiology to anatomical organization and transcriptomically defined cell types. Neuropixels recordings aligned to the Allen Common Coordinate Framework are combined with MERFISH spatial-transcriptomic atlases; channel-level and neuron-level features decode hierarchical brain-region identity and local transcriptomic composition, and MERFISH-aware generative models of waveform features show that regional waveform diversity is partly explained by local cell-type composition — providing a quantitative bridge between in vivo physiology and the cellular taxonomy of the brain. On the developmental axis, large-scale recordings in postnatal mouse cortex paired with EEG from preterm and term human infants identify a transient quiescent period preceding the emergence of mature cortical dynamics, suggesting a conserved milestone in the maturation of cortical circuits. In a clinical setting, the temporal coupling between interictal epileptiform discharges and physiologic sleep spindles is shown, in a pediatric focal-epilepsy cohort, to track cognitive deficits and predict post-surgical seizure outcome.

Taken together, these contributions argue that as recording technologies continue to expand the scale and resolution at which neural systems can be observed, computational approaches capable of integrating information across modalities, organisms, and levels of organization will be central to understanding both normal brain function and its dysfunction.