Tutorial 1: Dynamic System and Optimal Control Perspective of Deep Learning and Beyond

Date & Time

Wednesday, 14 Nov 2018

Tutorial Speaker

Bin Dong, Peking University, China

Abstract

Deep learning has achieved great success in many machine learning tasks and has major academic and industrial impacts. Deep architecture design has been one of the key topics in deep learning. Most of architecture designs are empirical and in lack of guiding principles. This tutorial will review some of the recent work on linking dynamic systems with deep architecture, and understanding deep neural network training as optimal control. It will show how we can take advantage of the rich knowledge in dynamic system and optimal control to provide guidance in designing new and effective deep architectures. On the other hand, such perspective also enables us to bring deep learning in applied mathematics to tackle challenging problems.

Tutorial 2: Dual Learning: Algorithms, Applications and Challenges

Date & Time

Wednesday, 14 Nov 2018

Tutorial Speaker

Tao Qin, Microsoft Research Asia, China

Abstract

While structural duality is common in AI, most learning algorithms have not exploited it in learning/inference. Dual learning is a new learning framework that leverages the primal-dual structure of AI tasks to obtain effective feedback or regularization signals to enhance the learning/inference process. Dual learning has been studied in different learning settings. This tutorial will first introduce several dual learning algorithms: (1) dual unsupervised learning, (2) dual supervised learning, (3) dual transfer learning, and (4) dual inference. Then it will cover multiple applications, including neural machine translation, image understanding, sentiment analysis, question answering/generation, image translation, etc. At the end, the tutorial will describe several challenges of dual learning, such as theoretical understanding, efficiency and scalability, and discuss future research directions.

Tutorial 3: Interpretable Anomaly Detection Cancelled

Date & Time

Wednesday, 14 Nov 2018

Tutorial Speakers

Evgeny Burnaev, Skolkovo Institute of Science and Technology, Russia

Dmitry Smolyakov, Skolkovo Institute of Science and Technology, Russia

Abstract

Anomalies are the unusual, unexpected, surprising patterns in the observed world. Identifying, understanding, and predicting anomalies from data form one of the key pillars of modern data mining. Effective detection of anomalies allows extracting critical information from data which can then be used for a variety of applications, such as to stop malicious intruders, detect and repair faults in complex systems, and better understand the behavior of natural, social, and engineered systems. Today, anomaly detection usually constitutes only a part of a bigger business process in real life applications. As such, we need not only to detect anomalies, but also to interpret our decisions in order to react properly. For example, for aircraft predictive maintenance we have to foresee failures, so that anomaly detection in measurements of different sensors is only the first step in solving this problem: after detecting an anomaly, an automatic analysis of such decision is needed to extract anomalous events leading to the failure of interest. In particular, such analysis should provide the decision-maker with actionable interpretations and information that could improve their reasoning on how to remedy the underlying issues leading to the failure. The stated interpretability problem is even more complicated due to complex and multi-modal nature of modern data such as web search logs, graphs, text, images, video, etc. For these types of data, model interpretation presents an unprecedented challenge, since in this case anomaly detection models are usually represented by complex and highly nonlinear functionals, ensembles of detectors, and often neural networks that are used to pre-process data and extract features. Therefore, the current tutorial proposes to highlight issues related to interpretable anomaly detection with additional emphasis on advanced anomaly detection methodologies including (but not limited to) those based on deep representation learning and ensembles.