SkipOOD: Efficient Out-of-Distribution Input Detection using Skipping Mechanism
Abstract: Detection of out-of-distribution (OOD) inputs has been a popular research direction in deep neural network (DNN) research, as OOD inputs can cause an undesirable decrease in the accuracy of the model. Furthermore, because of the deployment of DNNs on resource-constrained devices, it has become important to detect OOD inputs early during inference to save on computation.
In this work, we focus on detecting OOD inputs in a partial inference setting and investigate whether the skipping mechanism used in dynamic neural networks (DyNNs) can be leveraged for early OOD detection. We first establish that the feature maps at various DyNN gates can help identify anomalies. Building on this, we propose SkipOOD, a lightweight OOD detector that uses an uncertainty scoring function and an exit detector at each gate to robustly identify OODs as early as possible. Through extensive evaluation, we demonstrate that SkipOOD achieves competitive performance in detecting OOD samples while reducing resource usage by nearly 50%.
Submission Length: Regular submission (no more than 12 pages of main content)
Assigned Action Editor: ~Chuan_Guo1
Submission Number: 3810
Loading