Fundamental Research
Deep Learning
With access to extraordinarily large data, we focus on modeling the underlying structure for both structured and unstructured data. Our research covers fundamental and theoretical work in the field of deep learning, including mathematics, statistics, learning and algorithms. The most critical mission at IIAI is to advance in the understanding of underlying mechanisms, giving rise to intelligence. We will be dedicating more efforts on exploring high-level reasoning and thinking skills, that can eventually create human-imitative intelligence. These advances in the development of fundamental machine learning algorithm research and it application to real world scenarios will accelerate innovation and eventually benefit the society. Our primary research areas include but not limited to: deep neural network understanding, unsupervised learning (e.g. large-scale data unsupervised clustering, unsupervised/weakly-supervised data representation learning and unsupervised data knowledge association), adversarial learning (e.g. Generative Adversarial Networks and their derivatives).
Reinforcement Learning
Reinforcement Learning (RL) is an important area of machine learning inspired by behaviorist psychology, concerned with an agent learning how to behave in an environment by performing actions so as to maximize some notion of cumulative reward. Inception Institute focuses on Reinforcement Learning, including both policy-search and temporal-difference based, as well as closely related methods for sample-based decision theoretic planning. Our efforts will be fully dedicated to developing AI agents, which are able to learn in an interactive environment by exploring solutions and automatically determine the ideal behaviour. Fundamentally, the AI agents operate by getting the feedback information from trial and error using their own actions and experiences. For RL research, the techniques such as Monte Carlo tree search, swarm intelligence, control theory, operations research, information theory, optimisation theory and deep learning are expected to play key roles. Additionally, we apply RL in some real-word applications, including automatic natural language processing, speech analysis, healthcare, image processing and video analytics.
Computer Vision
Computer vision is an interdisciplinary field that seeks to build a human-imitative visual understanding system to allow machines to obtain high-level understanding from multiple visual data forms (e.g., images and video sequences). From a scientific perspective, we desperately promote the development of computer vision theories behind the artificial visual understanding system, and target at driving excellence in a wide range of unsolved fundamental computer vision problems, including but not limited to human action recognition, gait recognition, person re-identification. From a technological perspective, we heavily promote the transferability of our research findings for enhancing the performance of existing visual understanding systems.
Natural Language Processing
Language is an advanced and complex system of human communication. Over the past decades, Natural Language Process (NLP) has been a research area for developing machines’ capabilities in terms of how to use humans’ communication language. Based on the latest advancement in NLP theories, IIAI targets developing innovative text input and understanding solutions, as a universal and flexible interface that allows machines to intelligently and accurately understand voices and texts. In addition, IIAI also targets at developing highly effective and efficient NLP techniques that are capable of dealing with large volumes of natural language data in multiple documentation forms. Eventually, we aim to build an artificial language system that will be the interface of machines not only to humans, but also the whole textual and linguistic world.
Applied Research
Intelligent Video Understanding
We aim to build a world-leading intelligent video understanding platform that can unlock a new era of how to effectively utilize public and private monitoring and control devices with highly sophisticated video analytics in face detection and annotation, pedestrian identification and tracking, scene analysis and tagging, human behavior analysis and content enhancement. For example, with access to the national database that stores a vast amount of photo IDs, a fully-integrated person detection, tracking and recognition system can rapidly identify a target with deep learning techniques applied to a massive nationwide video system. Even more mission-critically, we can go from “unknown” to “unknown”, developing a self-learning AI system for detecting unknown human behavior from real-time surveillance camera feed, which can identify, classify and report abnormal activities in unlimited behavioral forms. Self-taught learning algorithms can study abnormal human patterns across the board and correspondingly perform inferences from the future patterns not yet known to us.
Medical Imaging
Medical image computing is an interdisciplinary field at the intersection of computer science, data science, mathematics and has a wide range of applications in medicine and bioscience. At the IIAI, researchers are fully dedicated to applying cutting-edge machine learning and computer vision techniques to revolutionize the healthcare industry through medical image computation and analysis, computer-assisted interventional systems and robotics, bioscience applications and computer-aided diagnosis and clinical visualization and feedback systems. All these applications require the manipulation and integration of medical image information. Many applications also depend on integration of the image information with sensor data (e.g. from tracking systems), and/or effector control systems (e.g. robots or positioning devices), and/or visual displays or other feedback systems. In fact, our researchers put efforts on a broad range of medical imaging applications. For instance, one of our key applications will be destined to Mammography, which is a medical imaging technique that uses a low-dose X-ray system to aid in the early detection and diagnosis of breast diseases in women. Advanced deep learning techniques can be applied much more accurately and efficiently to detect cancerous lesions than the traditional examination by expert radiologists. Another challenging target is Gastrointestinal Diseases (GI), where we can deploy AI-based medical imaging to attain highly effective detection and classification. Additionally, medical image computing techniques are utilized in eye fundus scan for diabetes early detection and cardiac disease diagnosis in our institute as well.
Healthcare
AI has extensive use in healthcare cases. We aim at redesigning healthcare by utilizing the latest advancement of AI research in drug development, computer-assisted interventional systems and robotics, bioscience applications and computer-aided diagnosis. For example, Artificial Intelligence is providing advice that could help more than 30 million Americans suffering from diabetes. The advice is coming from algorithms which are designed to help patients with diabetes to control their glucose levels by using data on how food, exercise, and insulin affect their glucose levels under different conditions. In addition, a care manager might use the office-based platform to ask for a trend analysis of when blood glucose levels for an individual patient have been high or low during certain days, weeks, or months. Those results might be used to adjust a patient’s overall care plan.