IC-1003Pretrained ResNet-50 and ViT-B/16 exhibit neuron activation patterns that are separable between in-distribution and out-of-distribution inputs, enabling post-hoc OOD detection without model modification
Yibing Liu, Chris XING TIAN, Haoliang Li, Lei Ma, Shiqi Wang
The paper formulates neuron activation states as the product of raw neuron outputs and KL-divergence gradients, then measures the coverage of these states under in-distribution training data. On pretrained ResNet-50 and ViT-B/16 (both trained on ImageNet-1k), out-of-distribution inputs trigger neuron states that fall outside the expected coverage area, producing lower coverage scores than in-distribution inputs. This separability in the models' internal neuron behavior allows a post-hoc uncertainty score (NAC-UE) to distinguish IN from OOD samples without modifying the model weights, achieving state-of-the-art AUROC across three OOD datasets.
The separability is measured using the paper's specific neuron state formulation (z ⊙ ∂dkl/∂z); raw neuron outputs alone show less separation (FPR95 35.72% vs 16.58% on layer4 of ResNet-50). The method requires a small subset of training data (1,000–50,000 images) to build the coverage model.