About Me

Weizhi Li, Gautam Dasarathy, and Visar Berisha. "Matched-Pair Experimental Design with Active Learning." Conference on Uncertainty in Artificial Intelligence (UAI), 2026. Statistical validityIntelligent data acquisition

Integrates machine learning with matched-pair experiments to detect causal treatment effects under limited resources. The proposed design uses agnostic active learning to prioritize treatment-responsive regions while identifying an enclosing region that provides an upper bound on the size of the responder region under the stated assumptions. This supports decisions in multiphase clinical trials: whether a treatment may benefit a sufficiently broad population to justify a larger, more expensive experiment.

Weizhi Li, Natalie Klein, Brendan Gifford, Elizabeth Sklute, Carey Legett, and Samuel Clegg. "Regularization via f-Divergence: An Application to Multi-Oxide Spectroscopic Analysis." JGR: Machine Learning and Computation, 2026. Adaptive learning

Develops a neural-network training approach for stable spectroscopic characterization of oxide weights under extreme simulated Martian conditions. The approach embeds comparisons between multi-oxide predictions and certified reference weights in an f-divergence loss, guiding the network to learn the underlying multi-oxide distributions. The left diagram shows how f-divergence regularization complements the prediction-error loss during training; the right illustrates large versus small divergence between predicted and reference oxide-weight distributions.

Pavel Dolin, Weizhi Li, Gautam Dasarathy, and Visar Berisha. "Statistically Valid Post-Deployment Monitoring Should Be Standard for AI-Based Digital Health." Conference on Neural Information Processing Systems (NeurIPS), 2025. Statistical validity

Healthcare operating conditions are dynamic: patient demographics, clinical practices, equipment, disease patterns, and data-acquisition procedures can shift. Yet a review cited in this paper found that only 9% of FDA-registered AI-based healthcare tools included a post-deployment surveillance plan. The paper argues for statistically valid, label-efficient post-deployment monitoring to detect shifts in data and model performance while controlling explicit error rates.

Weizhi Li, Visar Berisha, and Gautam Dasarathy. "Advanced Tutorial: Label-Efficient Two-Sample Tests." Winter Simulation Conference (WSC), 2024. Statistical validityIntelligent data acquisition

A tutorial on resource-efficient methods for revealing dependence on an exposure variable in industrial-scale simulation applications. Using the familiar machine-learning variables of features and labels, it presents statistical tools to detect whether labels depend on features before committing to large-scale computation, such as developing AI surrogate models for scientific simulations. These tools enable a pilot study to gauge whether data contain predictive signals and are ready for machine-learning modeling.

Weizhi Li, Prad Kadambi, Pouria Saidi, Karthikeyan Natesan Ramamurthy, Gautam Dasarathy, and Visar Berisha. "Active Sequential Two-Sample Testing." Transactions on Machine Learning Research (TMLR), 2024. Statistical validityIntelligent data acquisition

Proposes a framework for detecting observational dependence between measured features and an exposure or group variable when labels are costly. The design combines machine learning with active, sequential two-sample testing to decide where to collect data and monitor the accumulated evidence. Anytime-valid inference guides the next step: continue collecting data, stop without rejecting when the budget is exhausted, or reject independence to establish evidence of observational dependence.

Weizhi Li, Gautam Dasarathy, Karthikeyan Natesan Ramamurthy, and Visar Berisha. "A Label-Efficient Two-Sample Test." Conference on Uncertainty in Artificial Intelligence (UAI), 2022. Statistical validityIntelligent data acquisition

Introduces bimodal querying, which reverses the uncertainty-seeking principle of conventional active learning: in binary classification, it requests annotations where the model is most confident in class 0 or class 1. Combined with a two-sample test, this exposes predictive signals with fewer labels. Subsequent analysis establishes asymptotic maximization of feature-label mutual information under consistency conditions. The approach supports rapid signal assessment, including evaluating whether data contain useful predictive information for an AI task.

Weizhi Li, Gautam Dasarathy, Karthikeyan Natesan Ramamurthy, and Visar Berisha. "Finding the Homology of Decision Boundaries with Active Learning." Conference on Neural Information Processing Systems (NeurIPS), 2020. Adaptive learningIntelligent data acquisition

Proposes an active-learning framework to efficiently estimate the topological properties of a decision-boundary manifold. These properties support evaluating AI tools from a model marketplace using a topology-based comparison metric. The work establishes conditions under which boundary homology is recoverable, grounding the metric's reliability and the selection of models for downstream tasks. The left diagram illustrates the boundary geometry and sampling conditions; the right shows the active homology-estimation pipeline.

Weizhi Li, Gautam Dasarathy, and Visar Berisha. "Regularization via Structural Label Smoothing." International Conference on Artificial Intelligence and Statistics (AISTATS), 2020. Adaptive learning

Characterizes intrinsic structural complexity through local class overlap and develops structural label smoothing to adapt neural-network training to that complexity. By choosing smoothing strengths from the data's structure rather than imposing uniform regularization, the method adapts a general-purpose learning algorithm to a downstream task using its intrinsic structural properties. The figure traces the process from the original data through local complexity estimates to adaptive smoothing strengths.