Data Science Summer Hangouts Series 2026

Designing and Deploying AI in Public Systems: Evidence from Healthcare and Education

This talk discusses how machine learning and AI systems can be designed, deployed, and evaluated in real-world public systems. Drawing on case studies of medicine allocation in Sierra Leone and personalized AI tutoring in Taiwan, the talk covers the technical design of AI systems, including decision-aware machine learning and LLM-guided reinforcement learning, and the operational challenges of implementing and evaluating these systems through field deployment and experimentation.

Training Very Small Language Models on Device

In this workshop we will explore different ways to train and run small language models on personal computers using GPUs. We will follow a step-by-step tutorial to train models using the nanoGPT/nanochat frameworks. Then we will use an adaptation of the framework to train and visualize the training process on the browser and explore some of the capabilities of transformers.js (a library that supports running models in the browser using WebGPU).

Training on-device works best with a GPU with CUDA support or an Apple Silicon chip (you should be able to run things on any CPU, but it will be slow). You can still participate even if you don't have an accessible GPU, as some of the activities will be on the browser.

HackTheScan: Fine-Tuning Foundation Models for Clinical Imaging

Foundation vision models can be leveraged to extract features from clinical imaging data to assist in diagnosing disease more accurately. However, fine-tuning is required to apply these models to specific clinical tasks. In this hackathon-style workshop, attendees will work in teams to fine-tune MedSAM, a foundation vision model for medical imaging, in the context of Polycystic Ovary Syndrome (PCOS), recently renamed Polyendocrine Metabolic Ovarian Syndrome (PMOS). Participants will leave with hands-on experience adapting foundation models to real clinical workflows. Expect a lot of hands-on learning!

The Cloud is Just Someone Else's Computer

All of the information we enter into a computer (broadly defined) is represented somewhere physical. As we rapidly move toward “AI everything,” the concepts of where information resides, how it moves, and how it is processed are becoming increasingly obscured. This talk covers a mix of computing fundamentals and more abstract networking concepts that can help build intuition about what is actually happening at the level of data when we use modern AI systems. The goal is to provide a framework for thinking about the practical implications of AI in research, including questions of data governance, privacy, security, and trust.

Using Language Models as A Psychological Instrument

This tutorial will cover how language models can be used as a psychological instrument for measuring and analyzing human thoughts, beliefs, values, and decision processes. The session will begin with traditional psychological approaches for measuring human values and attitudes, including the use of surveys and rating scales, then explore how large language models (LLMs) can serve as a type of psychological instrument for analyzing natural language data. The tutorial will introduce practical approaches for using LLMs to code think-aloud protocols and open-ended survey responses.

Causal Thinking in Social Science

This session introduces how social scientists conceptualize causality and what it takes to make a credible causal claim. We cover the counterfactual logic of causal reasoning, the pitfalls when attempting to establishing causality and an overview of the toolkit for establishing causation, from experiments to design-based methods.