In progress
Human-in-the-loop evaluation of generative AI
Estimates human preference win rates across multiple evaluation metrics while minimizing the number of judgments needed.
COMPUTER VISION · AI · STATISTICS
Max Hamilton
Final-year PhD candidate at UMass Amherst. I combine AI predictions with human labels and statistical estimation to produce numbers we can trust.
RESEARCH DIRECTIONS
Selected work
01Reliable measurement with AI + human labelsUnbiased estimates and honest uncertainty from imperfect models3 papers · 1 ongoing
AI can process a collection quickly, but systematic errors can make its totals unreliable. This work uses models to guide a small amount of human labeling, then uses statistical estimators to correct the result and quantify uncertainty.
UAI 2026
Scalable Model-Assisted Multi-Target Estimation in Large Image Collections ↗Max Hamilton, Jinlin Lai, Daniel Sheldon, Subhransu Maji
Estimates many quantities at once, such as the relative abundance of multiple coral species, using AI model predictions and limited expert labels.
AISTATS 2026
Active Measurement of Two-Point Correlations ↗Max Hamilton, Daniel Sheldon, Subhransu Maji
Extends active measurement to spatial statistics, estimating how rare targets such as star clusters group across a galaxy.
NeurIPS 2025
Active Measurement: Efficient Estimation at Scale ↗Max Hamilton*, Jinlin Lai*, Wenlong Zhao, Subhransu Maji†, Daniel Sheldon† *equal contribution · †equal advising
Introduces an adaptive framework for human-in-the-loop measurement with unbiased estimators, confidence intervals, and sampling without replacement.
02Species distribution modelingLearning where species live from limited observations and other evidence2 papers
Species range maps are essential for ecology and conservation, yet most species have few recorded observations. These projects use shared representations, language, and images to make range estimation work with less data.
ICML 2025
Feedforward Few-shot Species Range Estimation ↗Christian Lange, Max Hamilton, Elijah Cole, Alexander Shepard, Samuel Heinrich, Angela Zhu, Subhransu Maji, Grant Van Horn, Oisin Mac Aodha
Predicts the range of a previously unseen species from a few observed locations, with optional text or image metadata, in a single feedforward pass.
NeurIPS 2024
Combining Observational Data and Language for Species Range Estimation ↗Max Hamilton, Christian Lange, Elijah Cole, Alexander Shepard, Samuel Heinrich, Oisin Mac Aodha, Grant Van Horn, Subhransu Maji
Connects citizen-science observations with Wikipedia descriptions to estimate ranges for species with little or no location data.
03Earlier computer vision workVideo prediction and large-scale 3D dataset curation2 papers
CVPR 2022
Comparing Correspondences: Video Prediction with Correspondence-wise Losses ↗Daniel Geng, Max Hamilton, Andrew Owens
Uses optical-flow correspondences to produce sharper, more perceptually accurate video predictions.
CVPR 2020
OASIS: A Large-Scale Dataset for Single Image 3D in the Wild ↗Weifeng Chen, Shengyi Qian, David Fan, Noriyuki Kojima, Max Hamilton, Jia Deng
A dataset of detailed 3D geometry annotations for 140,000 diverse real-world images.
BACKGROUND
Schools & research labs
2022—present
University of Massachusetts Amherst
PhD, Computer Science
- Computer Vision Lab
- Advisor: Subhransu Maji ↗
- Collaboration with Daniel Sheldon ↗
2016—2022
University of Michigan
MSE & BSE, Computer Science
Minor in Mathematics
- Owens Research Group · Andrew Owens ↗
- Fouhey AI Lab · David Fouhey ↗
- Vision & Learning Lab · Jia Deng ↗
INTERNSHIPS
SUMMER 2026
Adobe
PhD Research Intern · human-in-the-loop evaluation of generative AI models
SUMMER 2024
Dolby Laboratories
PhD Research Intern · efficient diffusion models for depth and surface-normal estimation
ACTIVE MEASUREMENT · INTERACTIVE EXPLANATION
How many birds are in the marsh?
1
The objective: estimate the total number of birds without inspecting every region.
2Your role: count the birds in each highlighted crop.
THE FIXED POPULATIONOne marsh · 48 regions
labeled current crop
SAMPLING STRATEGYROUND 1 / 48
or
Leave both unselected to alternate.
MAGNIFIED CROP MODEL PREDICTION: ——
How many birds are in this crop?
Choose a count to begin.
WHAT WE KNOW SO FAR
Estimated birds in the whole image
Uniform samplingEvery unseen region is equally likely
—estimated birds
0120240
Label 3 regions to see an approximate 95% confidence interval.
Model-guided sampling oursRegions are sampled by predicted bird count
—estimated birds
0120240
Label 3 regions to see an approximate 95% confidence interval.
UNBIASED regardless of AI prediction errorsVALID UNCERTAINTY confidence intervals account for sampling
How are samples and estimates calculated?
Model predictions. We use AI predictions to focus human review on the regions expected to be most informative. Crucially, the framework’s unbiasedness and confidence interval guarantees hold regardless of the accuracy or quality of those predictions. Better predictions lead to narrower confidence intervals.
Sampling. Uniform sampling ignores the model predictions and chooses equally among unseen regions. Our model-guided method samples without replacement using the underlying continuous model score; the count shown above is its rounded display value. A small exploration floor ensures every region can be selected, including those predicted as empty.
Estimation. Each model-guided draw uses an inverse-probability estimate of birds in the unseen image. Draw-level estimates are combined with fixed without-replacement weights from the Active Measurement paper. Intervals: uniform uses a finite-population variance and model-guided sampling uses the paper’s conditional variance construction.