The implicit bias phenomenon in deep learning

SPAAM Seminar Series: 19/02/2026 Title: The implicit bias phenomenon in deep learning Speaker: Yunus Cobanoglu, University of Warwick Abstract: In overparameterized deep learning, many networks can interpolate the data exactly, so the empirical loss has many global minimizers with zero training loss. This means that the optimization method (including choices such as initialization and learning rate) influences which minimizer is actually found, creating an implicit bias. This talk will cover how for simplified nonlinear models, gradient descent tends to select solutions of low complexity. Understanding this optimization-driven selection effect is important for understanding generalization, and it is still not fully understood.