0522204 20240103223735.420200214235959.9 9 s. E cav_un_epca*0522203VancouverUniversity of Oxford Computer Science department2019 approximation generative models orthogonal combinations cav_un_auth*0101207 Šmídl Václav UTIA-B Adaptivní systémy Department of Adaptive Systems AS AS Department of Adaptive Systems Ústav teorie informace a automatizace AV ČR, v. v. i. cav_un_auth*0389606 Bím J. CZ cav_un_auth*0307300 Pevný T. CZ http://library.utia.cas.cz/separaty/2020/AS/smidl-0522204.pdf GA18-21409S GA ČR cav_un_auth*0374053 This paper presents a new approximation of the marginal likelihood of generative models which is used as a score for anomaly detection. The score is motivated by the shortcoming of the popular reconstruction error that it can behave arbitrarily outside the known samples. The proposed score corrects this by orthogonal combination of the reconstruction error and the likelihood in the latent space. As experimentally shown on benchmark problems from anomaly detection and illustrated on a toy problem, this combination lends the score robustness to outliers. Generative models evaluated with this score outperformed the competing methods especially in tasks of learning distribution from data corrupted by anomalies. Finally, the score is compatible with contemporary generative models, namely variational auto-encoders and generative adversarial networks cav_un_auth*0389608 NeurIPS 2019 20191208 20191214 Vancouver CA BD 10000 10200 10201 2021 3 PO RVO:67985556 http://hdl.handle.net/11104/0308914 1 S 48 2019 SCOPUS PUBMED WOS cav_un_epca*0522203 Bayesian Deep Learning NeurIPS 2019 Workshop Vancouver University of Oxford Computer Science department 2019