Doctoral Thesis: Learning Possibilities: Stochastic Models for Vision

Candidate: Marianne Rakic

Title: Learning Possibilities: Stochastic Models for Vision

Abstract: When asking a question about visual data, there is often no single right answer. Input images can be blurry, low contrast, or otherwise difficult to interpret; different users hold different priors and biases; and the desired output frequently depends on the downstream task. Yet most machine learning models for vision produce a single deterministic prediction, collapsing this inherent variability and discarding information about the space of plausible solutions. Stochastic generative models, such as diffusion and flow matching models, help capture this variability by sampling multiple predictions from a learned distribution.
This thesis develop various strategies to produce diverse, plausible outputs across two different domains: medical imaging and photography. We first introduce Pancakes, a method that, given a medical image from a previously unseen domain, automatically generates multiple plausible whole-image segmentation maps consistently across images from the same domain. If the user knows the desired segmentation, they can provide a small context set of annotated image–segmentation pairs. In our new method, Tyche, we present an in-context learning framework that produces multiple plausible segmentations for this new task without retraining. Finally, in the context of photographic enhancement, we show how to transform everyday snapshots into images that look as if they were taken by a professional photographer, using text to disentangle the content of an image from its aesthetics and to sample multiple content-preserving enhancements.

Time: Apr 30, 2026 10:30 AM Eastern Time (US and Canada)

Join Zoom Meeting
https://mit.zoom.us/j/99775621099

Thesis Committee:
John Guttag, Adrian Dalca, Fredo Durand

Location: 32-G882 (Stata Center, Hewlett Room)

Details

• Date: Thursday, April 30
• Time: 10:30 am - 11:30 am
• Category: Thesis Defense
• Location: 32-G882 (Stata Center, Hewlett Room)