Based Loss

Based loss functions optimize models by directly focusing on the ranking of predictions, rather than solely on individual prediction scores, aligning training with evaluation metrics like average precision. Current research emphasizes improving the efficiency of these losses, particularly for large datasets, through techniques like bucketing and developing differentiable surrogates for non-differentiable ranking operators. This approach shows promise across diverse applications, including object detection, edge detection, image retrieval, and text ranking, leading to improved performance and addressing challenges like class imbalance and label uncertainty.

Papers