TIME-LAPSE: Learning to say “I don't know” through spatio-temporal uncertainty scoring

Safe deployment of trained ML models requires determining when input samples go out-of-distribution (OOD) and refraining from making uncertain predictions on them. Existing approaches inspect test samples in isolation to estimate their corresponding predictive uncertainty. However, in the real-world, deployed models typically see test inputs consecutively and predict labels continuously over time during inference. In this work, we propose TIME-LAPSE, a spatio-temporal framework for uncertainty scoring that examines the sequence of predictions prior to the current sample to determine its predictive uncertainty. Our key insight is that in-distribution samples will be more “similar” to each other compared to OOD samples, not just over the encoding latent-space but also across time. Specifically, (a) our spatial uncertainty score estimates how different OOD latent-space representations are from those of an in-distribution set using metrics such as Mahalanobis distance and cosine similarity and (b) our temporal uncertainty score determines deviations in correlations over time using representations of past inputs in a non-parametric, sliding-window based algorithm. We evaluate TIME-LAPSE on both audio and vision tasks using public datasets and further benchmark our approach on a challenging, real-world, electroencephalograms (EEG) dataset for seizure detection. We achieve state-of-the-art results for OOD detection in the audio and EEG domain and observe considerable gains in semantically corrected vision benchmarks. We show that TIME-LAPSE is more driven by semantic content compared to other methods, i.e., it is more robust to dataset statistics. We also propose a sequential OOD detection evaluation framework to emulate real-life drift settings and show that TIME-LAPSE outperforms spatial methods significantly.