Learning Oriented Computer Networks

点击翻译按钮获取中文摘要

The "data pipe" model used by the existing Internet protocol stack is no longer adequate for many emerging applications. Information, which can have different data representations, and knowledge, e.g., learned model parameters, have communication patterns and requirements different from those of point-to-point, reliable and elastic data transfer. Information and knowledge can be aggregated and transformed inside the network, due to latency constraints, and control scope is no longer limited to endpoint-to-endpoint only. On the other hand, new communication infrastructures such as 6G mobile communication systems, low-orbit satellites and underwater/ground communication systems are emerging rapidly, which offer additional communication capabilities with certain topological features. Now the network protocol stack between application requirements and communication services becomes a new bottleneck. This research program focuses on a new computer network architecture serving learning-centric applications in data centers, around network edges and on mobile devices, by leveraging, improving and creating new learning techniques to determine and optimize protocol mechanisms and control policies. It will further our research in the last two decades on topology control, protocol design and performance evaluation for computer networks, to network a truly dependable cyber-infrastructure. The learning-oriented computer network can learn from applications and communications automatically and continuously while running on different infrastructures to support diverse requirements, and keep evolving its protocol mechanisms and control policies in an online manner, while maintaining security and preserving privacy, to learn and perform more effectively and efficiently. Specifically, using decomposition, protocol stack can be decomposed into functional building blocks to reduce redundancy and repetition in the current layered protocol stack. The learning and sequential decision-making algorithms can compose certain building blocks for a given application and configure their operating parameters automatically based on the learned application requirements and network dynamics. This new approach can fundamentally change the existing approach of limited functionality and performance in the current, fixed protocol stack, and open new space for future applications. On the other hand, the modular computer network architecture will support in-network information processing with computing, storage and control for decentralized, federated and distributed learning applications. Through this research program, a diverse and inclusive team of undergrad and graduate students, postdoctoral fellows and industry personnel will be trained and worked with, and the research outcomes can directly benefit the Canadian society and economy by leveraging the Canadian leadership in artificial intelligence to meet the need for dependable, high-performance and low-cost cyber-infrastructures.