CVPR Daily - Tuesday

15 DAILY CVPR Tuesday Jiawei hopes this work will encourage researchers to understand feature distribution better when performing any task. It is essential because any operation should be mathematically analyzed and clearly understood as to why it is helpful for the task. In this paper, he analyzes inter-class separation and intra-class compactness to discover why certain aspects of the conventional framework are essential. He would encourage researchers to focus on bias theory in conventional training, especially deep learning . “ We always encounter long-tail datasets , and this imbalance can be interpreted differently, ” he adds. “ It can be the number of training data, it can also be the times that this environment happens, and we would like to encourage the researchers to focus on the bias toward this understanding and then use the analysis on the feature geometry, which is also called the distribution of features, to develop their projects. ” To learn more about Jiawei’s work, visit Poster 305 this morning from 10:30-12:30 in the West Exhibit Hall. DiGeo: Discriminative Geometry-Aware Learning ...