Self-Anomaly-Detection Model Training via Initialized Meta Model

Anomaly detection has become a key challenge affecting the training accuracy of machine learning. Because the training data is usually collected from Internet, many noised samples will be captured and these samples can decrease the model training accuracy. However, because the abnormal samples are difficult to predict when the samples are collected, and the training samples collected may contain many unknown exception categories, and the labels of normal samples may be incorrect, in this case, it is difficult to train an anomaly detection model based on supervised learning to accurately identify the anomaly samples. In this paper, we propose a new unsupervised anomaly detection method based on BiGAN, namely Rt-BiGAN, to identify the outliers in the training data. Firstly, we propose a Bigan network initialization method based on meta-learning algorithm with a small number of normal samples. Then, a self-supervised drop training is designed to improve the detection ability of the model. Finally, we evaluate our Rt-BiGAN over real-world datasets and the simulations results demonstrate that our mechanism is effective to detect the outliers in model training data.

I. Introduction

With the development of big data, many computer vision jobs require enormous training datasets to achieve the deep learning model training, which is a form of data hunger [1]. To meet the data requirements, model trainer usually picks up the training data from the Internet to achieve classification and other visual research [2], [3], [4], [5]. In a typical Internet picture-driven application, we can look for photos on web search engines like Google and Bing or on photo-sharing sites, download a large number of images that correlate to a text query, and then model the target object/concept linked with the image collection. Search engine or site crawler images, on the other hand, usually captures some noised samples, which can impact the learning model. As a result, the outliers, or irrelevant images, in the collected training data need to be removed.