Lei Song
Principal Researcher at Microsoft Research Asia
Dr Lei Song, Principal Researcher at Microsoft Research Asia, has many years of research and practical experience in the field of artificial intelligence and industrial optimisation, focusing on the application of artificial intelligence and optimisation algorithms in the industry, especially in the fields and directions of supply chain, energy, and sustainable development. He holds several patents in related fields and has published dozens of papers in international conferences and journals. Prior to joining Microsoft, Song Lei worked in an e-commerce company responsible for the development of algorithms and system construction of intelligent supply chain system, and accumulated a lot of experience in related business areas. Currently, his main research focus is on the application of reinforcement learning in large-scale operations research optimisation problems, and he is also interested in the application of basic models in industrial scenarios, especially in knowledge management and decision support.
Topic
Context-sensitive decision optimisation: an exploration of the application of fundamental models to industrial control
Recent developments in Fundamental Models (FMs) have shown that FMs trained on large corpora have acquired knowledge in many domains, greatly improving the effectiveness of traditional natural language processing tasks. Traditionally, Foundation Models have been used mainly in static environments for natural language tasks; however, their potential application in dynamic industrial control environments, especially in changing environments, has not yet been fully demonstrated. This talk will present a novel approach to enable base models to adapt to changing environments in control scenarios without extensive fine-tuning through a context learning (ICL) algorithm. The ICL algorithm proposed by Microsoft Research Asia introduces a trainable agent model that is able to retrieve historical samples in a context-aware manner at the right time, thus enabling the base model's decision-making ability in control tasks. By strategically selecting relevant samples at each step, the base model demonstrates better performance in dynamically changing environments.
