Machine Learning
Algorithm
We work on challenging machine learning problems and design cutting-edge algorithms for both real-world applications and fundamental science problems. In particular, we are interested in deep learning, reinforcement learning, neural architecture search, pre-training, causal learning, etc.
Theory
Machine learning, with deep neural networks, has achieved breakthroughs in AI tasks and impressive progress in science. However, due to the non-convexity of neural networks, machine learning is suffering from high computation and sample cost. In addition, it has high variance and is hard to generalize to new domains. We take the theoretical view to understand and improve the technologies in machine learning taxonomy, in terms of sample/computation efficiency, robustness, interpretability, privacy-preservation, etc.
Application
In recent years, machine learning algorithms (especially deep learning and reinforcement learning) have greatly boosted the performance of many real-world applications, and applications have also driven the progress of machine learning research (e.g., ResNet originated from image classification, Transformer from machine translation). We work on multiple key applications, including language and speech, music, game testing, finance, logistics, healthcare (including drug discovery and precision medicine).
AI for Science
Physics
Machine learning is disrupting physics research. We are using machine learning, especially deep learning, to tackle physics problems that are extremely challenging to solve before. The laws of nature are described as partial differential equations (PDEs). With AI techniques, we can leverage big data to solve, simulate, or predict known PDEs more efficiently and discover unknown laws. Specifically, we take machine learning approaches to solve parametric PDEs through physics-based self-supervision, to predict the stationary state of PDEs from simulated data, to detect parameters of PDEs by modeling their symmetry groups, and to speed up molecular dynamics simulation.
Biology
Biology comes to the big data era. At Microsoft Research Asia, we seek to unlock the big biological data and reveal the secret of life with computation. We are working on advanced techniques for computational molecular biology to help scientists obtain deep insights into life at the molecular level. We are carrying out research in the following areas: molecular dynamics simulation, genomics, immunomics, microbiome, and protein folding.
Sustainability
Sustainability involves a few systematic and scientific challenges, including energy crisis, global warming, air pollution, and other thorny issues. Most of them essentially lie in over-reliance on fossil fuels and are accompanied by high carbon emissions. To accelerate the energy transition and facilitate carbon removal, we need computational physics to develop new technologies. For example, density functional theory is the foundation to calculate the properties of crystals so to aid material discovery, and molecular dynamics plays the basis to accelerate the adsorption and conversion of carbon dioxide. We work on these core techniques to benefit sustainability research.
AI for Industry
Healthcare
Healthcare plays an important role in human beings. To understand the human body and better care human health, we pay great efforts on the research of biology understanding, drug discovery, and medical forecasting. The biology sequence data, e.g., DNA, RNA, protein, is fundamental to the life system. Drug discovery with AI technique hugely increases the speed of drug innovation. Medical forecasting with accurate and robust predictions can greatly support clinical decision-making and treatment. We exploit the machine learning power in these fields to benefit healthcare research.
Finance
The financial industry has adopted statistical analysis for different tasks for a long time and has accumulated tremendous valuable data. These conditions leave a big potential for AI technologies to empower the financial industry. In particular, we start with the intelligent quant investment as our first exploration area. Now we also expand our research on RegTech like anti-money laundry. We mainly focus on several typical challenges/research directions in applying AI techniques to Machine learning. We also build an open-source AI-oriented quant investment platform Qlib to accelerate the research exploration and algorithm landing.
Supply Chain
In supply chain and logistics systems, we need to solve complex decision-making problems to maximize revenue while minimizing operational costs, for instance, to optimize inventory and vehicle routing. At Microsoft Research Asia, we focus on developing efficient Deep Reinforcement Learning algorithms for optimization problems in logistics. We are targeting a scalable and generalizable DRL approach that can be applied to solving problems in practice, hence, to empower the companies and customers in the logistic industry.