With 5G coming to reality, the next question is, what will 6G be? Mobile cellular standards are updated every decade, and its evolution is always driven by new application demands and user scenarios in mobile networks. From 1G to 5G, the application scenarios have evolved from phone calls, text messages, to online browsing, video streaming, and online gaming.
For 6G that is ahead of us, we expect that information data used for Artificial Intelligence (AI) tasks will possibly bypass human-perceived applications and become the dominant source of mobile data traffic. The related applications include autonomous vehicles, industrial automation, virtual reality/augmented reality, and E-health systems. Naturally, this means that next-generation wireless networks need to evolve to cope with the communication requirements of AI-oriented data traffic.
From centralized to decentralized Machine Learning systems
Conventionally, Machine Learning (ML) algorithms are performed in a centralized unit where the training data sets are stored. In many applications, the training data are collected from different user devices and those data might contain private information about the users. Due to the privacy concerns in centralized ML systems, decentralized ML has emerged as an alternative solution that allows user devices to perform intelligent information processing locally based on their raw data or observations, without uploading them to a centralized server/cloud.
For any distributed system that relies on signaling and information exchange between distributed nodes to achieve some collective goals, wireless connectivity will always be the performance bottleneck. In a wireless network, the amount of information bits that can be reliably delivered is limited by the available communication resources (time, frequency, space). An efficient communication protocol and resource allocation design for distributed ML systems can make a significant difference than blindly applying the conventional rate-driven wireless design.
Research Activities
Our current research directions in this project are:
- Resource allocation for federated edge learning
- Over-the-Air (OtA) computation for distributed learning and estimation
- Communication-efficient designs for distributed consensus and optimization.
The specific topics cover data compression, resource allocation, signal processing, medium access control, privacy, and security aspects of distributed intelligence over wireless networks. Broadly speaking, our research group aims to identify new topics and interesting problems within the intersection of communication theory, distributed algorithms, and machine learning.
Acknowledgements: our research group is supported by ELLIIT, Swedish Research Council (VR), Wallenberg Foundations, and Zenith.
