Spotlight paper at NeurIPS

Our paper on recalibration of classifers was selected as a Spotlight paper at NeurIPS this year. NeurIPS is the most popular and competitive conference in AI and only 3% of submissions were selected for Spotlight this year.  The paper proposes a new method with theoretical guarantees for adapting pretrained neural networks to a new domain not well represented in the pre-training set. The method is directly applicable the in-context learning problem of foundation models (large language models).

Z Sun, D Song and A Hero, “Minimum-Risk Recalibration of Classifiers,” Neural Information Processing Symposium (NeurIPS), New Orleans, to appear, Dec 2023. https://arxiv.org/abs/2305.10886. Spotlight. NSF IPA, ARO MSU, DOE ETI.