微软亚洲研究院 (Audio) - Channel 9

Hopcroft Interview

Microsoft Research Asia Research Groups

AI research at Microsoft Research Asia will be introduced, including machine learning, computer vision, natural language processing, knowledge mining and urban computing based on big data. Specifically, I will share our research on developing new learning algorithms and distributed machine learning platform for training very big models on big data based on heterogeneous hardware (e.g. CPU, GPU and FPGA cluster). In addition to deep learning, we are also working on knowledge mining and symbolic learning that integrates facts, common sense, and logic rules in a unified knowledge representation for machine comprehension of text. I will introduce these research works and show how they have been used to build artificially intelligent and socially engaging conversational agents such as XiaoIce and Microsoft's Cognitive Services. 

Future Talent 2040

Given the investment and evidence of progress in Artificial Intelligence some suggest that it is merely a matter of time until AI can match, complement or surpass human intelligence. This session looks at recent research advances in machine learning and cognitive science and discusses the needs and design principles to support fundamental research in AI. Together we will look at how to push AI technology towards more natural human-AI communication and interaction that will facilitate social learning and collaborations between humans and AI agents.

In this session, we will discuss machine learning from two extremes, theory and application. The goal is to bring theory and application researchers together and inspire each other's work, so that the theory research can become more targeted and the application research can have better theoretical guarantees.

In recent years, the growing research on deep learning and reinforcement learning has led to many exciting breakthroughs. However, on the other hand, there are still many open problems regarding deep learning and reinforcement learning. In this session, we will reflect, problem-solve and discuss key missing elements, as well as synthesizing the opportunities for academia and industry to further advance this field.

In recent years, the growing research on social multimedia and visual Q&A has led to many exciting breakthroughs. However, on the other hand, there are still many open problems regarding social multimedia and visual Q&A. In this session, we are going to do some reflection on this important research field, and discuss what's missing and what are the opportunities for academia and industry to further advance this field.

In recent years, the growing research on deep learning has led to many exciting breakthroughs in vision and multimedia communities. However, on the other hand, there are still many open problems regarding deep learning for vision and multimedia. In this session, we are going to do some reflection on this important research field, and discuss what's missing and what are the opportunities for academia and industry to further advance this field.

It is evident that when machine learning meets big data, it is vital to have a powerful system/infrastructure to support the distributed training task. In recent years, people have used different frameworks for this purpose, including iterative MapReduce, parameter server, and data flow. In this session, we are going to discuss how to enhance these frameworks from both system and algorithmic perspectives, and how to implement parallel machine learning algorithms under these frameworks. In addition, we will discuss the future trend of machine learning system and infrastructure and how to push its frontier through close collaboration between academia and industry.

Urban computing connects ubiquitous sensing technologies, advanced data management & machine learning models, and novel visualization methods to create win-win-win solutions that improve urban environment, human life quality, and city operation systems.

Capturing and reconstruction 3D real world (scene, objects, and dynamic human characters) plays important role in VR/AR applications. However, real time high quality 3D capturing and reconstruction is still a very challenging task. In this session, we are going to do some reflection on this important research field, and discuss what's missing and what are the opportunities for academia and industry to further advance this field.

This session examines the future direction of robotics research. As a background movement, AI is sparking great interest and exploration. In order to realize AI in human society, it is necessary to embody such AI in physical forms, namely to have physical forms. Under such circumstance, this session explores and clarifies the current direction of basic robotics research. Thorough examination of what types of research components are missing, and how does such capability development affect the directional paths of research will be highlighted. 

Generative learning and discriminative learning are two major approaches in machine learning. The fast development of deep learning demonstrates the power of discriminative learning. In recent years, some have started to integrate generative learning into the deep learning process in order to incorporate prior knowledge. In this session, we will discuss the pros and cons of each approach and how to seamlessly integrate them together.

Gaining an in-depth understanding of users is critical for building artificial intelligence systems. With the rapid development of positioning, sensing and social networking technologies, large quantities of human behavioral data are now readily available. They reflect various aspects of human mobility and activities in the physical world. The availability of this data presents an unprecedented opportunity to user understanding. In addition, recent studies in psychology suggest that numerous psychological features, such as personality traits, are highly correlated to user behaviors. It will be interesting to study how we can design computational frameworks for inferring psychological features of users, based on their data at different levels and across heterogeneous domains, and how these frameworks can benefit the development of artificial intelligence systems. In this session, we plan to invite researchers from computer science, psychology and cognitive science areas. We will brainstorm innovative ideas, technologies, systems and applications along this interdisciplinary research direction.

微软计算思维

How can we educate the world's population in a scalable, affordable way?  This question is driving fascinating research at the intersection of human-computer interaction, social computing, natural language processing, machine learning, and learning sciences.    I'll discuss the state-of-the-art in what is becoming known as learning at scale, with a focus on how to improve peer feedback, how to automate grading, and how to help instructors understand what the students understand.  I will emphasize how tackling this problem is leading to new socio-technical innovations. 

Panel – Engaging with AI: How Human and Machine can Work Together to Shape the Future

We are living in a world that is witnessing an incredible pace of innovation. Despite all the rapid advancements – or perhaps, because of them – top companies are working incredibly hard to find new ideas that will create new business opportunities. In other words, companies are more desperate than ever to find the Next Big Thing that will shake up the industry and create growth. What is remarkable today is that, in this quest, companies are increasingly turning to advanced research. This talk will outline the excitement in the industry that surrounds these eventful times, with a focus on the exhilarating journey that researchers are experiencing at Microsoft. 

Cyber-security today is focused largely on defending against known attacks. We learn about the latest attack and find a patch to defend against it. Our defenses thus improve only after they have been successfully penetrated. This is a recipe to ensure some attackers succeed—not a recipe for achieving system trustworthiness. We must move beyond reacting to yesterday's attacks and instead start building systems whose trustworthiness derives from first principles. Yet, today we lack such a science base for cybersecurity. That science of security would have to include attacks, defense mechanisms, and security properties; its laws would characterize how these relate. This talk will discuss examples of such laws and suggest avenues for future exploration.