We propose a method to enable elastic inference for pretrained ViTs via structured pruning. We do not require labels, generalize to models without a classification head and do not need re-training. We improve over the existing state of the art.

DISCOVR is a self-supervised framework for echocardiography video representation learning that integrates spatial and temporal modeling, achieving strong generalization in anomaly detection, segmentation, and LVEF prediction.

We systematically collect and unify different notions of 'linear regions' of ReLU networks from the literature and study the computational complexity of counting them.