Tuesday 28 October 2025 Host: Henry GoukSpeaker: Henry Gouk, Philipp Bomatter, and Jack GearyTitle: Deep Learning with Distribution ShiftsAbstract: We train models, we deploy them, and everything breaks. Mismatches in the data distributions that models are exposed to during training and after deployment are ubiquitous in the so-called "real world". In this talk, Henry and his students will discuss some of our recent efforts to address common distribution shift patterns. Henry will kick things off with some high-level theoretical ideas related to how the problem can be formulated and some fundamental limits we must contend with. One real-world example of distribution shifts are participant distribution shifts in electroencephalogram (EEG) data. Philipp will present his recent work investigating this problem through the lens of data scaling studies, controlling both overall sample size and participant diversity. A particularly pathological class of distribution shifts are those that arise due to data sources responding to the deployment of a specific model. Jack will discuss the weaknesses and limitations associated with current approaches for simulating such responses, and present their recent work that uses a Lagrangian-based approach to overcome these limitations and produce responses that better approximate the theoretically expected strategic behaviour. Oct 28 2025 13.00 - 14.00 Tuesday 28 October 2025 Speaker: Henry Gouk, Philipp Bomatter, Jack Geary IF G.03
Tuesday 28 October 2025 Host: Henry GoukSpeaker: Henry Gouk, Philipp Bomatter, and Jack GearyTitle: Deep Learning with Distribution ShiftsAbstract: We train models, we deploy them, and everything breaks. Mismatches in the data distributions that models are exposed to during training and after deployment are ubiquitous in the so-called "real world". In this talk, Henry and his students will discuss some of our recent efforts to address common distribution shift patterns. Henry will kick things off with some high-level theoretical ideas related to how the problem can be formulated and some fundamental limits we must contend with. One real-world example of distribution shifts are participant distribution shifts in electroencephalogram (EEG) data. Philipp will present his recent work investigating this problem through the lens of data scaling studies, controlling both overall sample size and participant diversity. A particularly pathological class of distribution shifts are those that arise due to data sources responding to the deployment of a specific model. Jack will discuss the weaknesses and limitations associated with current approaches for simulating such responses, and present their recent work that uses a Lagrangian-based approach to overcome these limitations and produce responses that better approximate the theoretically expected strategic behaviour. Oct 28 2025 13.00 - 14.00 Tuesday 28 October 2025 Speaker: Henry Gouk, Philipp Bomatter, Jack Geary IF G.03
