Search papers, labs, and topics across Lattice.
This paper tackles the problem of unreliable likelihoods for OOD detection in deep generative models by exploiting the properties of continuous normalizing flows. They show that OOD samples are mapped to atypical noise samples under the noise prior, an effect not captured by likelihoods alone. Based on this, they propose Signal in the Noise (SITN), a method that detects OOD samples by identifying these atypical noise mappings, achieving strong performance without OOD data access or significant overhead.
Forget likelihoods: OOD samples become detectable "signal" when you look at how normalizing flows map them into the latent noise space.
Deep generative models offer a natural foundation for out-of-distribution (OOD) detection, yet prior work has shown that their assigned likelihoods are notoriously unreliable indicators for in- vs out-of-distribution data. In this paper, we address this problem by leveraging the diffeomorphic and mass-preserving properties of continuous normalising flows. Our analysis shows that OOD samples are mapped to noise samples that are highly atypical under the noise prior in ways not captured by the likelihood. Based on this observation, we propose a new method -- Signal in the Noise (SITN) -- for OOD detection on the single-sample level. SITN requires no access to OOD data, incurs minimal computational overhead, and provides strict control of false positive rates. Comprehensive evaluations through standard benchmarks and synthetic perturbations highlight the method's effectiveness and the absence of the complexity bias inherent to likelihood-based methods.
