ML+Coffee: How Can I Apply ML/AI To My Data?

ML+Coffee offers a supportive environment to discuss ongoing ML/AI projects and share knowledge across campus. Whether you're looking for advice on applying ML/AI to your data, hoping to demo a tool or method, or interested in discussing a paper, ML+Coffee offers the perfect space. The majority of our attendees are applied practitioners from diverse fields, not AI/ML purists looking to critique. Coffee and pastries provided ☕, courtesy of our sponsors.

Where: Room 1145, Discovery Building (330 N. Orchard St.).
When: Monthly on Wednesdays, 9-11am CT. Fall 2026 dates include: 9/9, 10/7, 11/11, 12/9.
Register: RSVP for ML+Coffee

Oct 7 Schedule

9:00-9:10 - Intros 

9:10-10:00 - Interpreting Neural Net Features with Sparse Autoencoders (Siwei Qian)
I'd like to lead a discussion of two papers, Toy Models of Superposition (Elhage et al., 2022) and Towards Monosemanticity (Bricken et al., 2023). I found them hard to follow, so the plan is to work through the core ideas together - what a "feature" means here, what the superposition hypothesis says, and why it makes individual neurons hard to interpret. Then we'll look at how the second paper uses a sparse autoencoder to pull features back out, and why simpler methods like SVD fall short when features are in superposition. No interpretability background needed.

10:00-11:00 - Detecting Insect Migration in Weather Radar Data (Yi Liu)
I'm part of a research lab studying insect migration, particularly using weather radar data together with ground observations. Weather radars generate large datasets consisting of repeated 3-D volume scans of the atmosphere. One of our major challenges is distinguishing insect signals from other signals, such as precipitation, birds, and bats. One direction I'm particularly interested in is taking advantage of the 3-D spatial structure of insect migration. Migrating insects often concentrate within relatively thin layers at particular altitudes, which could provide useful information for classification. However, radar observations are collected in polar coordinates and only sample portions of the atmosphere. A conventional approach would be to project and interpolate the radar observations onto a regular 3-D Cartesian grid before applying image- or volume-based machine-learning methods. At continental scales and over many years of radar observations, however, this can become extremely computationally expensive. I'm therefore wondering if there is a more efficient way to use the native 3-D radar observations for machine learning without first interpolating every scan onto a regular 3-D grid. For example, could we design a model that directly works with the radar's polar geometry or irregularly sampled 3-D data? Ultimately, our goal is to develop a model that can automatically detect and characterize insect migration around individual radars, and then connect detections across radar stations and through time to reconstruct large-scale insect migration pathways. If you have experience with ML for irregular spatial data, 3-D object detection, radar data, point clouds, graph neural networks, or related problems, I would love to hear your ideas and recommendations during ML+Coffee!

Have a demo, paper, or ongoing project to discuss? Add your name to the discussion queue. No formal presentation is required - this event prioritizes open dialogue and casual discussion. If helpful, bring a couple of slides or visuals (e.g., to share data, methods, or results). Many participants just bring a few key points or a rough overview of their work, and the conversation flows naturally from there.