MOOD - Multi-level 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:

Lin, Ziqian et al. “MOOD: Multi-level Out-of-distribution Detection.” 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2021): 15308-15318.

Annotations


  • Existing OODdetection methods commonly rely on a scoring function thatderives statistics from the penultimate layer or output layerof the neural network * show annotation

  • existing solutions require a full feedforward pass for anygiven test-time input and utilize a fixed amount of computa-tion * show annotation

  • computational cost of OOD detection can be ex-acerbated by the over-parameterization of neural networks,which nowadays have reached unprecedented depth and ca-pacity * show annotation

  • how canwe enable out-of-distribution detection that can adjust andsave computations adaptively on-the-fly * show annotation

  • adaptive OOD detection frameworkbased on intermediate classifier outputs * show annotation

  • exploits intermedi-ate classifier outputs for dynamic and efficient OOD infer-ence * show annotation

  • direct relationship be-tween the OOD data complexity and optimal exit level, andshow that easy OOD examples can be effectively detectedearly without propagating to deeper layers. * show annotation