Malgorzata Schwab and Ashis Biswas, University of Colorado, USA

In this paper we explore the applicability of Invertible Neural Network architecture for anomaly detection techniques on time series data and hypothesize that a reversible network designed with embedded convolutional transformations is an excellent fit for that task. We leverage previous findings on autoencoders as well as deep generative maximum-likelihood training focused primarily on processing images and apply them in the innovative way to the time series data exemplified by electrocardiograms or industrial sensor data. We recognize a challenge of common denominator patterns that occur across the entire sample domain, which might dominate the likelihoods and introduce intrinsic bias. We then mitigate it by applying wavelet transforms to decompose a time series into a set of subcomponents to eliminate low-level similarities between the healthy and abnormal samples. We conclude that the Invertible Neural Network designed to solve inverse problems learns data reconstructions extremely well, and thus provides a remarkable solution for anomaly detection that is applicable to medical diagnostics, as well as other use cases in the similar problem space, such as predictive maintenance or detecting out-of-distribution inputs to protect integrity of systems relying on machine learning components.

Invertible, Autoencoder, Anomaly.y