An approach to anomaly detection is to use a partly disentangled representation of the latent space of a generative model. In this study, generative adversarial networks (GAN) are used as the normal data generator, and an additional encoder is trained to map data to the latent space. Then, a data anomaly can be detected by a reconstruction error and a position in the latent space. If the latent space is disentangled (in a sense that some latent variables are interpretable and can characterize the data), the anomaly is also characterized by the mapped position in the latent space. The present study proposes a method to characterize temporal anomalies in time series using Causal InfoGAN, proposed by Kurutach et al., to disentangle a state space of the dynamics of time-series data. Temporal anomalies are quantified by the transitions in the acquired state space. The proposed method is applied to four-dimensional biological dataset: morphological data of a genetically manipulated embryo. Computer experiments are conducted on three-dimensional data of the cell (nuclear) division dynamics in early embryonic development of C. elegans, which lead to the detection of morphological and temporal anomalies caused by the knockdown of lethal genes.
CITATION STYLE
Ueda, T., Tohsato, Y., & Nishikawa, I. (2020). Temporal Anomaly Detection by Deep Generative Models with Applications to Biological Data. In Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics) (Vol. 12396 LNCS, pp. 553–565). Springer Science and Business Media Deutschland GmbH. https://doi.org/10.1007/978-3-030-61609-0_44
Mendeley helps you to discover research relevant for your work.