<ul><li>Introducing the first one-stage Top-k Learning-to-Defer framework, merging prediction and deferral by training a joint score-based model.</li><li>The approach optimizes prediction and deferral across multiple entities simultaneously through a single end-to-end objective.</li><li>The framework includes a cost-sensitive loss and a novel convex surrogate that generalizes across Top-k regimes without the need for retraining.</li><li>Experimental results on CIFAR-10 and SVHN datasets demonstrate the superiority of the one-stage Top-k method over traditional Top-1 deferral and the effectiveness of the adaptive variant, Top-k(x), in balancing predictive accuracy and consultation cost.</li></ul>

One-Stage Top-$k$ Learning-to-Defer: Score-Based Surrogates with Theoretical Guarantees

Discover more