Research portfolio

• Per-channel artifact and epileptiform-transient rejection, detrending, line-noise removal and bandpass filtering within the RNS bandwidth.
• Theta-band (~4–12 Hz) instantaneous amplitude and phase from time-frequency decomposition, with oscillatory bouts quantified against a fitted 1/f background.
• Synchronized neural, behavioral and positional streams. Circular–linear statistics relate theta to distance-to-target and spatial location.

• Spatial-cue front end built from binaural ITD/ILD signals feeds excitatory–inhibitory cortical units, so attentional gain acts on spatially indexed channels.
• Evaluated with spike-timing-based measures, not firing rate alone: how decodable the target stays as competitors are added across the spatial grid.
• A route toward biologically grounded sound segregation for hearing-assistive devices in multi-talker environments.

• Grounded in the predictive coding theory of cortical function.
• Learns hierarchical timescales through event demarcation and extracts variable-length temporal representations from a continuous observation stream.

Agents in the open world see non-stationary, temporally correlated streams. Standard SGD assumes i.i.d. samples from a stationary distribution. Co-designed with Shivang Patel, this framework learns quickly and stays stable on data with temporal dependencies.

• Efficient trackers overfit to in-distribution data. SiamABC improves accuracy markedly on out-of-distribution sequences.
• A forward-only, backward-free update rule on batch-norm statistics adapts to the incoming stream at negligible latency.
• CPU-efficient template–search attention and new self-supervised losses offset the limited capacity of efficient trackers.

Retraining a segmentation model for every new microscopy collection costs annotation and compute. With Voke Brume, I designed contrastive losses that adapt to the target domain with pseudo-labels while staying anchored to the source. The result is clearly better segmentation of non-convex cell instances from only a few labeled samples.

• Co-designed with Matthew Keaton. Instead of chasing one fully general model, the method adapts to novel data from a handful of annotations.
• Extended to 3D volumes and integrated into syGlass for VR microscopy quantification, shown at SfN 2022 and 2023.
• A follow-up contrastive representation loss earned a Trainee Highlight Award at the 9th Annual BRAIN Initiative Meeting.

• AutoTrack: efficient domain adaptation for single-particle tracking in 3D time-series microscopy, with a spatiotemporal fusion module framed as attention.
• syFind: 3D cell quantification and segmentation that adapts its representations to unseen microscopy data automatically.
• Done in the Vision & Learning Group under the lab's contract with IstoVisio. Integrated with their front-end team into the shipping syGlass product and demoed at the SfN 2022 Expo.

• Continuous multivariate sensor telemetry in Seeq, cleaned for irregular sampling, sensor drift and missing segments.
• A VAE-based LSTM learns a latent model of nominal machine behavior and flags deviations through reconstruction and prediction residuals, with no labeled fault data needed.
• ONNX models deployed to the Seeq/Sanofi cloud for live predictive-maintenance decisions.

A receptive-field-based shunting network with neurophysiologically valid dynamics and Hebbian-like learning for object recognition. Predictive coding emerged from the formulation, and a formal derivation produced a neurophysiologically plausible training procedure for deep networks.

A biologically plausible Bayesian learning framework for continual adaptability and reasoning under uncertainty. An adaptive thresholding mechanism acts as a variational activation that mimics leaky integrate-and-fire dynamics, and it was extended to spiking neural networks for neuromorphic hardware.

A test-time adaptation method for cellular instance segmentation on top of Cellpose. New losses inspired by Shannon entropy and self-supervised contrastive learning improve performance without additional labels or access to the source data.

A bottom-up approach that detects plant organs (leaf, flower, fruit, bark, high-density leaves) and fuses the predictions of a variable number of organ-based species classifiers. This counters data variability and long-tail effects. Presented at WVU 2020, URDC 2021 and NCUR 2021.

Autonomous collection of wearable physiological time series, with cleaning, resampling and windowed feature extraction for irregularly sampled streams, and models trained to infer target health outcomes. Part of the NSF Artificial Intelligence and Health 2020 workshop.