Can AI models really be too confident?
Stefano Giampiccolo’s PhD work shows how to make AI models smarter and more trustworthy.
He’s developing a new ensemble method for Neural and Universal ODEs that reduces overconfidence in AI models for dynamical systems. By using mode connectivity to keep ensemble members diverse, his approach improves reliability in reconstructing both fully and partially data-driven models, tested on systems like Lotka–Volterra, damped oscillator, and the Lorenz attractor.
The COSBI team discussed his research internally yesterday — with ideas flowing as freely as the coffee brewed the Italian way, straight from the moka pot.
Big congratulations to Stefano Giampiccolo for sharing his work with us and pushing the boundaries of uncertainty quantification!
