Unsupervised Deep Learning Framework for Solving One Class Classification Problems

单类分类（OCC）是一种特殊的机器学习范式，在训练阶段只有来自正类/正常类的样本可用，而负样本要么完全不存在，要么采样不足，要么以未标记数据的形式存在。在正常数据很容易收集、标记或大量可用的情况下，OCC的重要性就变得迫在眉睫，而负面数据要么不可用，要么很少出现，要么难以标记，要么获得成本太高，要么收集这些数据可能造成健康和安全危害。OCC范式在各个领域都有应用，包括机器故障诊断、异常健康模式、欺诈检测、视频中的暴力检测、异常环境事件、罕见疾病检测、音频监控中看不见的行为以及癌细胞检测等等。在这些应用程序中，负面数据可能在训练阶段不可用，但它可能在测试阶段出现。检测其发生是非常重要的，因为这可能对健康、安全、经济和环境产生不利影响。

One-class classification (OCC) is a special machine learning paradigm where only the samples from the positive/normal class are available during the training phase, while the negative samples are either completely absent, poorly sampled or present in the form of unlabeled data. The importance of OCC becomes imminent in situations where the normal data is easy to collect, label or is available in abundance, whereas the negative data is either unavailable, occurs rarely, hard to label, too costly to obtain or its collection may cause health and safety hazards. The OCC paradigm finds applications in various fields, including machine fault diagnosis, abnormal health patterns, fraud detection, violence detection in videos, unusual environment events , rare disease detection, unseen behaviours in audio monitoring, and cancer cell detection, to name a few. In these applications, the negative data may not be available during training phase, yet it can occur during the testing phase. It is very important to detect its occurrence because that can adversely impact the health, safety, economics, and environment. A major problem in traditional OCC methods is that they require domain specific feature extraction to be performed on the raw data, which is not only ad-hoc but a tedious, time-consuming, and error-prone process. Feature extraction becomes even more challenging in OCC in the absence of negative data. Several neural networks based solutions exist to handle OCC problems; however, there is a clear absence of a deep learning framework for solving these problems. Therefore, I envisage that the most promising future direction in the research of OCC paradigm rests largely with developing innovative unsupervised deep learning solutions that facilitates automatic feature learning. In this discovery grant, my long term goal is to develop a research program that involves the development of an unsupervised deep learning framework for solving OCC problems. To achieve this goal successfully, following are my specific objectives: 1. Develop novel unsupervised deep learning algorithms and their ensemble to provide performance enhancements. 2. Handle heterogeneous data and develop adaptive multi-modal unsupervised deep learning methods to provide robust performance. 3. Create new solutions for positive-unlabeled learning problem using deep learning methods. During my graduate studies and postdoctoral research, I extensively worked in the field of OCC. This discovery research grant is a natural continuation of my research work in this field, which is demonstrated through my publication record. This grant has a high impact in terms of Developing novel unsupervised deep learning algorithms and network architectures, Addressing real world problems that can adversely impact the health, safety, economics, and environment, Disseminating knowledge by sharing datasets, code and algorithms, and Training six high quality HQPs to become future leaders of the field.