The paper ”COMPOSITE FEDERATED LEARNING WITH HETEROGENEOUS DATA” by Jiaojiao Zhang, Jiang Hu and Mikael Johansson has been awarded the ICASSP2024 Best Paper Award. This prestigious recognition was awarded at the 2024 IEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP 2024) held in Seoul, Korea, from April 14 to 19, 2024. ICASSP 2024 is renowned as the most influential conference in signal processing, featuring 2826 accepted papers this year.
Jiaojiao Zhang, who is the first author of the paper, is a Digital Futures postdoc with Mikael Johansson and Joakim Jaldén as faculty sponsors. The co-author Jiang Hu from Harvard University also contributed to the paper’s success. The ICASSP organizing committee commended the paper for its remarkable contribution to the field, showcasing innovation and academic excellence in acoustics, speech, and signal processing.
The authors propose a novel algorithm for solving composite Federated Learning (FL) problems in the paper. This algorithm strategically decouples the proximal operator and communication to manage non-smooth regularization. It also addresses client drift without making assumptions about data similarity. Moreover, each worker employs local updates to reduce communication frequency with the server, transmitting only a d-dimensional vector per communication round. The authors demonstrate the superiority of their algorithm over state-of-the-art methods through numerical experiments and prove its convergence linearly to a neighborhood of the optimal solution.
– I am deeply honoured to receive the ICASSP2024 Best Paper Award. This recognition is a testament to our research team’s hard work and dedication. I am grateful for the opportunity to contribute to the field of signal processing and look forward to continuing our innovative work in the future, says Jiaojiao. The position as Digital Futures postdoc fellow has truly been a great experience.
The award was presented to the winners during the Closing Ceremony of ICASSP2024 on Friday, April 19. According to Young-Cheol Park, ICASSP 2024 Awards Chair, this recognition reflects the significant impact of the recipients on the field, inspiring further motivation in their ongoing research endeavours. John Hansen, IEEE ICASSP-24 Tech Chair, emphasized the special distinction of being selected as the ”THE” Best Paper award recipient among over 3000 papers included in the IEEE ICASSP-24 Technical Program.
– I am confident that we will see more and more applications of federated learning in the near future, continues Jiaojiao. The benefits of using more data to train better models are very large. However, the design of federated learning algorithms is challenging. On the one hand, we must understand how to design algorithms that produce high-quality models without wasting too much time or energy on communication; our ICASSP paper contributes to this direction. On the other hand, we have to guarantee data privacy and earn users’ trust so that different organizations dare to collaborate. This is another area that I am active in and where there is a lot more to do. I also think that federated learning could be extended to solve many important problems that current algorithms cannot handle. Our ICASSP paper introduces several techniques that could be used to address much more complex problems than we have considered so far.
ICASSP is the world’s largest and most comprehensive technical conference focused on signal processing and its applications. It attracts thousands of professionals annually, offering a comprehensive technical program presenting the latest developments in research and technology in the industry. The conference venue, COEX, situated in the Gangnam District of Seoul, is renowned for being the birthplace of the famous ”Gangnam Style” music. It is a hub of technology, business, and culture, offering access to various unique Korean cultural experiences and delightful cuisines amidst the warm spring cherry blossoms near the Han River.
Link to the awarded paper: COMPOSITE FEDERATED LEARNING WITH HETEROGENEOUS DATA
Text: Johanna Gavefalk
