Towards Unsupervised Deep Learning Based Anomaly Detection

Novelty or anomaly detection is a challenging problem in many research disciplines without a general solution. In machine learning, inputs unlike the training data need to be identified. In areas where research involves taking measurements, identifying errant measurements is often necessary and occasionally vital. When monitoring the status of a system, some observations may indicate a potential system failure is occurring or may occur in the near future. The challenge is to identify the anomalous measurements that are usually sparse in comparison to the valid measurements. This paper presents a land-water classification problem as an anomaly detection problem to demonstrate the inability of a classifier to detect anomalies. A second problem requiring the identification of anomalous data uses a deep neural network (DNN) to perform a nonlinear regression as a method for the estimation of the probability that a given input is valid and not anomalous. A discussion of autoencoders is then proposed as an alternative to the supervised classification and regression approaches in an effort to remove the necessity of representing the anomalies in the training dataset.