BruinML · Research · July–October 2026
Finding unfamiliar regions in images.CLIP-Based Out-of-Distribution Detection
I worked with a five-person BruinML team on finding small out-of-distribution regions in images using frozen CLIP patch features.
- Group
- BruinML
- My work
- Simulation & evaluation
- Research period
- July–October 2026
The research question
A small unfamiliar region can disappear when an image is summarized as a whole. Our five-person BruinML team studied training-free detection of these regions using frozen CLIP patch features.
My experiments
I designed three Python Monte Carlo studies. The initial sweep covered 225 configurations: five noise distributions, five object geometries, and nine signal strengths. I compared global, maximum, top-k, and spatial-scan aggregation.
I kept fitting, calibration, and evaluation samples separate, recorded the configurations and results, and used Wilson intervals and paired bootstrap comparisons to examine uncertainty.
A clipped spatial scan
I implemented clipping and null-centering of patch scores to limit the effect of extreme values. In one Gaussian-mixture simulation, the clipped scan increased detection power from 19.9% to 97.7% relative to the original unclipped scan, at a 5% false-positive target.
That is a controlled simulation result for a particular setting, not an overall accuracy score or a measured improvement on every real-image dataset.
Geometry and dispersion
Fixed-size geometry and fixed-total-signal experiments helped isolate where aggregation fails. Spreading the same total signal across an image can make a scan less effective; it is a diagnostic finding rather than an improvement claim.
The manuscript
I produced three figures and reproducible experiment records and contributed experiments, analysis, and writing to the coauthored manuscript. We submitted it to Algorithmic Learning Theory (ALT 2027); it remains under review.
The CLIP project is completed. My ongoing BruinML work on demonstration ranking is a separate project.
