Back
The paper 'Efficient Optimization for Rank-based Loss Functions', on which Michal Rolinek was contributing to, received the CVPR Honorable Mention award.
In the paper 'Efficient Optimization for Rank-based Loss Functions', we give an efficient algorithm for computing a gradient of rank-based loss functions (such as average precision or NDCG). In practice, it means that optimizing the 'correct composite loss' (for example in object detection) no longer comes with a higher training time.
Arxiv link: https://arxiv.org/abs/1604.08269