Why Does Generative Learning Help Deterministic Prediction?

Generative Representation Learning Across Biological Modalities

Predicting molecular measurements from histology is often evaluated as a deterministic regression problem. Yet generative objectives such as conditional flow matching can still produce better representations and predictions even when stochastic sampling is not the goal.

The Question

What does a generative objective teach the pathology encoder about the molecular modality? Rather than treating performance gains as a black-box effect, this project studies how intermediate molecular states shape information routing and representation learning.

Approach

I use controlled interventions, representation analysis, and attention-based diagnostics to examine when target-side molecular information improves pretrained pathology features—and when a simpler deterministic objective should be sufficient.

Broader Lesson

Generative modeling can be valuable not only for producing samples, but also as a mechanism for transferring structure between modalities. Understanding that mechanism can lead to simpler and more reliable multimodal learning systems.

Publication links and figures will be added after the complete public metadata is confirmed.