Darius Knowledge Hub

Neural Mean Discrepancy for Efficient Out-of-Distribution Detection

1 min read

Neural Mean Discrepancy for Efficient Out-of-Distribution Detection

Created: =dateformat(this.file.ctime,"dd MMM yyyy, hh:mm a") | Modified: =dateformat(this.file.mtime,"dd MMM yyyy, hh:mm a") Tags:

Dong, Xin et al. “Neural Mean Discrepancy for Efficient Out-of-Distribution Detection.” 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2021): 19195-19205.

Annotations

  • hypothesis thatstandard off-the-shelf models may already contain sufficientinformation about the training set distribution which can beleveraged for reliable OOD detection * show annotation

  • propose a novel metric called Neural MeanDiscrepancy (NMD), which compares neural means of theinput examples and training data * show annotation

  • believe the off-the-shelf model itself should contain sufficient informationabout the training data distribution. So we proposed a sim-ple study (Figure 3) by looking at the model activation’smean for OOD and ID input batches. The result revealsthat the activation means of OOD mini-batches consistentlyand clearly deviate more from those of the training data. * show annotation

Graph View