Research highlights
My research focuses on efficient language model systems, reliable AI, reinforcement learning, and graph representation learning. I study how to adapt models with fewer resources, assess when their predictions can be trusted, and improve decision-making through feedback and relationships in data. Alongside these topics, I work on computer vision, medical imaging, and cloud computing. Some of this work reflects my interest in AI for Social Good, particularly agriculture, healthcare, and environmental sensing.
Adaptation & compression
I study how to adapt and simplify language models with limited training and computation. One line of work asks when learned input vectors, called prefixes, can reproduce a low-rank adapter's effect inside a frozen attention head. Another examines which model weights to remove so a text classifier can still return a useful set of possible labels. Under the calibration assumptions, that set has a statistical guarantee on how often it contains the correct answer.
Questions I study
Uses sensitivity to plausible labels to choose which weights to remove. Studies how to keep prediction sets small after pruning, with coverage supplied by independent calibration, and checks how the benefits change with recovery training and equal time budgets.
Identifies when a shared input prefix can reproduce a LoRA adapter's effect in a single frozen attention head under resource limits. Separates what is mathematically possible from what numerical precision allows; it does not establish whole-model equivalence.
Evidence & guarantees
I study what evidence is needed to trust a model's predictions or approve an update. This includes checking whether an agent's memory contains enough information to answer a question, deriving statistical bounds on adapted models' prediction errors, and deciding how earlier evaluations can help when fresh human labels are expensive. Across these problems, I examine the assumptions behind a guarantee and when the available evidence falls short.
Questions I study
Shows how a benchmark can accidentally reveal missing evidence through memory length. Introduces size-matched examples and studies an estimator of whether the remaining memory supports an answer, including where it fails to generalize.
Distinguishes when old evaluator errors can support a new release decision from when they can only guide fresh checks. Introduces a sequential test that can use helpful history while retaining validity when that history is unreliable.
Studies when a distribution over low-rank adapters supports a useful bound on prediction error. Identifies limits that better complexity accounting cannot fix, and shows how to tighten part of a bound without changing predictions.
Learning & decisions
I study how agents learn to choose actions from feedback while respecting costs and constraints. My work asks when an agent should pay for language-model advice, how to learn which choices work best while guaranteeing a minimum share of exposure, and how to compare states without checking every pair. For example, a recommendation system may need to give each provider a minimum number of opportunities to be shown. Fair exposure in recommendation is one way this research connects to AI for Social Good. Related work on graph-structured bandits appears in the graph representation learning section.
Questions I study
Studies a controller that predicts whether advice is worth requesting. Compares selective querying with simple schedules at matched budgets, showing why fewer calls alone do not establish better timing or a reliable deployment guarantee.
Turns fractional allocation plans into actual choices that meet minimum-exposure requirements in every time block. Derives bounds on reward lost relative to the best fair allocation, with additional conditions for overlapping groups.
Approximates how similarly states behave by updating selected pairs instead of every pair. Under structural assumptions, reduces pairwise computation and returns lower and upper bounds that show when the approximation is too uncertain to use.
Structure & representations
I explore how relationships between entities can help a model learn. In knowledge graphs, this means learning vector representations of entities and their relationships to predict missing links. In contextual bandits, it means using connections among available choices to build a smaller representation for learning from rewards. My work examines which parts of the graph help and when its structure can be misleading.
Questions I study
Trains a generative sampler to produce challenging, varied counterexamples for knowledge graph embeddings. Uses training-graph structure to reduce conflicts with missing true links, while recognizing that an unobserved link is not necessarily false.
Uses graph structure to build a smaller feature space for choosing actions from feedback. Derives learning guarantees that account for imperfect structure and provides a diagnostic for when graph-based reduction is likely to help.
Environment & agriculture
I build lightweight deep learning models for real-world vision problems under limited compute, from estimating local air quality with smartphone photos to identifying plant leaf diseases.
Questions I study
Estimates a location-specific aggregated air quality index from smartphone-captured images with an efficient deep convolutional network.
A limited-resource deep convolutional network for identifying plant leaf diseases.
Healthcare AI
I work on medical image analysis through small architectural changes, such as a single involution layer before a CNN that improves classification and segmentation with a negligible number of extra parameters, along with polyp segmentation and eye-tracking-based autism classification.
Questions I study
Adds a single involution layer before a CNN to improve medical image classification and segmentation with a negligible number of extra parameters.
Combines involution and convolution to classify eye-tracking patterns of children with autism spectrum disorder.
A bi-axial Mamba bottleneck for segmenting gastrointestinal polyps.
Distributed systems
I also work on cloud systems for image storage, retrieval, and content-based search in object storage. A concern across these projects is efficient deployment in environments with limited resources.
Questions I study
Orchestrates image retrieval and storage across a cloud system.
A content-based searching approach for object storage systems.
A connected agenda
Across these areas, I ask what a learning system can preserve when computation, labeled data, or opportunities to explore are limited. Efficient adaptation and graph representations change what a model needs to learn. Reliability methods check what its evidence supports, while sequential decision-making makes the costs of acting and gathering more information explicit.
See the complete publication record, including papers beyond these selected projects.
View all publications