Event Description

Modern deep learning delivers strong accuracy but says little about how much to trust any single prediction – a real obstacle in risk-sensitive, regulated settings. Classical kernel methods, by contrast, come with a mature theory of error, but have not, until recently, scaled to or plugged into deep architectures.

This talk presents a way to close that gap: kernel components that slot into standard deep-learning pipelines – from a classification head to a transformer sub-block – while carrying a computable, geometry-aware error estimate for each prediction, one that propagates through the full model. The result is a kernel-based counterpart to conformal prediction: a step toward deep models whose outputs are auditable, prediction by prediction, rather than a black box.

Across vision, reinforcement learning and tabular data, these kernel components match the accuracy of their neural counterparts – error-awareness need not cost accuracy. I will discuss what this opens up for finance and insurance, where quantifying and auditing model error is increasingly a requirement, not a luxury.