Time and date: 4 February 2026 at 2:00 pm | Location: Abacws 0.04 | Speaker: Thomas Greatrix
Ensemble methods are ubiquitous in modern AI, yet theoretical understanding of why they work is restricted to specific losses or simple averaging techniques. In this talk, we first introduce the mathematical foundations of a generalized ambiguity decomposition that decouples the loss function from the ensembling method, and show it to be stable across datasets and loss functions. We then demonstrate that BERT’s internal combiners can actually be outperformed by simple soft voting during overfitting, and reveal how diversity harms performance in non-convex regimes.
