Most machine learning models are trained on populations, making them accurate "on average" but unreliable for individuals.

In this session, we’ll explore how to bridge this gap using generative deep learning: pretraining a model on population data and fine-tuning it with an individual’s own data via transfer learning, without starting from scratch. We’ll also discuss why generating a range of plausible outcomes is often much more honest and useful than predicting a single number when dealing with real-world uncertainty.

Live Demo: We’ll see this approach applied to healthcare by building a personalized "digital twin" that simulates blood glucose trajectories in Type 1 Diabetes, showing how to test "what-if" counterfactual scenarios safely before applying them in real life.

Important! Bring your own laptop

Note: While we use diabetes as a concrete case study, this transfer learning pattern applies to any domain with shared data across users.

We’ll wrap up the session with some snacks and networking. We can't wait to see you there!