Research Vision
How can we learn invariant and generalizable representations that remain stable across domains, subjects, sensors, and time?
Neural signals (EEG and intracranial EEG) serve as a particularly demanding testbed for this problem — if models can generalize under the extreme heterogeneity of brain data, they can generalize in medical, industrial, and cyber-physical systems more broadly. This work sits at the intersection of representation learning, robust and generalizable AI, signal processing and systems theory, and human-machine systems.
Core Research Thrusts
1. Invariant Representation Learning Under Heterogeneity
Models trained on one population or environment often fail when deployed in another because learned representations become entangled with nuisance factors — subject differences, sensor variability, acquisition protocols — rather than task-relevant structure. The approach embeds invariance-enforcing objectives directly into representation learning, encouraging extraction of stable latent structure while suppressing domain-specific variation, and promoting cross-subject and cross-site generalization. These methods are architecture-agnostic and have been validated across CNNs, recurrent models, and transformers.
A core finding is that robust generalization requires explicitly modeling heterogeneity — naïve scaling often amplifies artifacts rather than improving invariance. These ideas extend naturally to wearable sensing, medical time-series, multimodal systems, and industrial monitoring.
2. Scalable Deep Learning for Clinical Neural Data
Clinical iEEG datasets present extreme variability: non-standard electrode layouts, limited labeled samples, site-specific acquisition differences, and strict data-sharing constraints. This line of work develops frameworks that carefully distinguish within-site performance from true cross-site generalization, and artifact-driven improvements from biologically meaningful learning.
The emphasis is on heterogeneity-aware pretraining, principled regularization, and evaluation protocols that reflect real deployment constraints rather than idealized benchmarks. The goal is AI systems that are clinically meaningful, not merely statistically impressive.
3. Robust Machine Learning for Closed-Loop Systems
Human-in-the-loop systems such as brain-computer interfaces create dynamic feedback loops between model predictions and user adaptation. Offline accuracy alone does not guarantee stability, safety, or reliable long-term interaction. This work integrates representation learning with control-theoretic concepts to analyze stability under distribution shift, robustness to artifacts, and error amplification in closed-loop settings — bridging machine learning, control systems, and human factors engineering for safety-critical real-time applications.
Future Directions
Theory-grounded invariant representation learning — formalizing generalization under heterogeneity and non-stationarity, moving beyond empirical validation toward principled guarantees.
Foundation-style models for temporal biomedical data — large-scale pretraining strategies tailored to the constraints of clinical and sensor data, where labels are scarce and acquisition protocols vary.
Robust evaluation paradigms — benchmarks that measure real-world deployment performance under domain shift, replacing leaderboard metrics that reward in-distribution fitting.
Safe and responsible neurotechnology — ethical, interpretable, and clinically validated AI systems for deployment in high-stakes environments.
The long-term goal is a research program centered on robust, deployable AI for heterogeneous real-world environments — grounded in neural systems, extending to broader cyber-physical and industrial domains.