Confidence-Aware Paced-Curriculum Learning by Label Smoothing for Surgical Scene Understanding

Curriculum learning and self-paced are the training strategies that gradually feed samples from easy to more complex. They have captivated increasing attention due their excellent performance in robotic vision. Most recent works focus on designing curricula based difficulty levels input or smoothing feature maps. However, labels control utility a curriculum manner is still unexplored. In this work, we design paced by label (P-CBLS) using with uniform (ULS) for classification tasks fuse spatially varying (SVLS) semantic segmentation manner. ULS SVLS, bigger factor value enforces heavy penalty true limits less information. Therefore, (CBLS). We set at beginning of decreased it zero model lower higher. also designed confidence-aware pacing function combined our CBLS investigate benefits various curricula. The proposed techniques validated four surgery datasets multi-class, multi-label classification, captioning, tasks. robustness method corrupting validation data into different severity levels. Our extensive analysis shows improves prediction accuracy robustness. code publicly available https://github.com/XuMengyaAmy/P-CBLS. Note Practitioners —The motivation article improve deep neural networks safety-critical applications such as controlling ability allowing imitate cognitive process humans animals. approaches do not add parameters require additional computational resources.