Â鶹´«Ã½AV

The 2024 Hong Kong, Guangzhou and Taipei Joint Workshop on Artificial Intelligence, Communications and Information Theory (AICIT 2024) took place at City University of Hong Kong (CityUHK) from July 20 to 21, co-organized by the Department of Computer Science at City University of Hong Kong and Â鶹´«Ã½AV Information Theory Society Hong Kong, Guangzhou, and Taipei Chapters. The workshop aims to provide a fruitful platform for AI and IT Scholars across the strait to exchange ideas and build collaborations and friendships. The Steering Committee consists of Prof. Raymond Wai Ho Yeung from the Chinese University of Hong Kong and Prof. Xiaohua Jia from City University of Hong Kong. The workshop was generously sponsored by Huawei Technology Co., Ltd. and the Institute of Network Coding (INC) at the Chinese University of Hong Kong. AICIT 2024 attracted over 70 researchers and students to participate, including the Chinese Institute of Electronics Information Theory Chapter chairman, Prof. Baoming Bai, and the vice chairman, Prof. Xiao Ma.

This year, AICIT 2024 featured 12 talks and 14 student posters covering a range of topics, including information theory, coding theory, wireless communications, machine learning, generative AI, etc. The General Chair of AICIT 2024 and Â鶹´«Ã½AV Information Theory Society Hong Kong Chapter Chair, Prof. Linqi Song, hosted the opening ceremony and welcomed all participants. The General Chair and the Â鶹´«Ã½AV Information Theory Society Guangzhou Chapter, Prof. Li Chen gave opening remarks and the background of the workshop. Prof. Chia-Han Lee, on behalf of the Â鶹´«Ã½AV Information Theory Society Taipei Chapter, expressed their best wishes for a successful workshop. The Â鶹´«Ã½AV Hong Kong Section Chair and the City University of Hong Kong Associate Provost, Prof. Ray Chak-Chung Cheung, delivered an opening remark online, expressing the Â鶹´«Ã½AV Hong Kong Section’s great support for the workshop and best wishes for the workshop’s success. Dr. Bo Bai from Huawei Technology Co., Ltd. warmly welcomed the attendees and expressed their support to the university-industrial collaborations. 

On Day 1, the Shannon Awardee, Prof. Raymond W. Yeung from the Chinese University of Hong Kong delivered a keynote speech about proving information inequalities by Gaussian elimination. Proving information inequalities and identities under linear constraints on the information measures is an important problem in information theory. For this purpose, ITIP and other variant algorithms have been developed and implemented, all of which are based on solving a linear program (LP). Prof. Raymond W. Yeung presented their symbolic computation approach which can solve such LPs very efficiently. In some cases, only the Gaussian elimination is needed without having to solve an LP. In other cases, the original LP is reduced to a much smaller LP. 

Besides the keynote speech, on Day 1, Prof. Cheuk Ting Li from the Chinese University of Hong Kong gave a talk on one-shot coding using Poisson processes. He proposed a Poisson functional representation approach for proving finite blocklength coding theorems in multiuser settings. This technique uses a query-based mechanism for encoding/decoding and bounds error probability via the Poisson matching lemma. Dr. Qiaosheng Zhang from Shanghai Artificial Intelligence Laboratory introduced an information-theory-inspired method, called information-directed sampling (IDS). IDS balances information gain (exploration) and immediate reward (exploitation). Prof. Shih-Chun Lin from National Taiwan University discussed the proposed optimal finite-length linear codes for broadcast packet erasure channels with feedback. He derived the exact optimal second-order achievability among all possible LNCs, closing the gap between state-of-the-art LNC second-order inner and outer bounds. Prof. Shenghui Song from The Hong Kong University of Science and Technology shared their results regarding the fundamental limits of several large-scale MIMO channels, including the non-Gaussian fading channels, the two-hop channels, and the two-hop wiretap channels. The last talk of Day 1 was delivered by Prof. Chia-Han Lee from National Yang Ming Chiao Tung University. He presented their recent progress in utilizing generative models for the physical layer design of wireless communication systems. He discussed GANs, VAEs, and diffusion models, and demonstrated their applications in federated learning, semantic communication, and joint source-channel coding for image data.

At the end of the first day, organizing committees prepared a banquet for speakers and attendants. During the dinner, the Chair of the Chinese Institute of Electronics Information Theory Chapter, Prof. Baoming Bai, gave a welcoming speech. He expressed the chapter’s encouragement and support for researchers from Guangzhou, Hong Kong, and Taipei to engage in academic exchanges.

Prof. Li Chen and Prof. Cheuk Ting Li chaired the Day 2 sessions. On the morning of Day 2, Prof. En-Hui Yang from the University of Waterloo delivered an online keynote speech first. He discussed how to enable information theory to jump on the bandwagon of deep learning-based artificial intelligence. Several insights were provided in this talk: (1) conditional mutual information (CMI) can be used to measure the concentration of a classification deep neural network (DNN) in the output probability distribution space of the DNN, and (2) optimization techniques in rate distortion function and channel capacity can be modified to minimize (maximize, resp.) CMI along with minimizing the standard cross entropy function in DL, yielding CMI constrained deep learning (CMIC-DL), knowledge distillation (KD) resistant DL, and KD-amplifying DL. 

Following the keynote speech from Prof. En-Hui Yang, Prof. Antoni B. Chan from City University of Hong Kong gave a talk about the next-generation Explainable AI framework that promotes human-AI mutual understanding. He highlighted the need for cognitive science-based approaches to enhance transparency and trust in AI, especially in critical systems. Prof. Yu-Chih Huang from National Yang Ming Chiao Tung University discussed the piecewise-stationary bandit problem from a minimalist perspective. This work introduced diminishing exploration, a mechanism that operates without knowing change points and improves the performance and regret scaling of existing algorithms. In the afternoon, Prof. Wenrui Dai from Shanghai Jiao Tong University introduced their recent work on rate-distortion optimization for learned image compression. Key points include using reversible autoencoders for stable transforms, analytic models for precise rate control, and generative entropy modeling for accurate rate estimation. Dr. Xueyan Niu from Huawei Technology Co., Ltd. gave a talk on exploring the strengths and weaknesses of Transformers through the lens of information theory. Moreover, she also showed her recent work on extending the context window of LLMs through the lens of information theory. The final talk was given by Prof. Qianqian Yang, who came from Zhejiang University. She showcased her work on semantics-oriented communications, using generative models to reduce transmitted information by over 99%. She also explored a system with evolving performance through caching and a method enhancing efficiency via probabilistic graphical models. 

After all the talks, AICIT 2024 also included a student poster session. 14 posters were displayed at the entrance of the venue. During the meeting break, participants also visited the CityU Indra and Harry Banga Gallery.

 At the end of AICIT 2024, Prof. Linqi Song and Prof. Li Chen hosted the closing ceremony. They concluded the workshop and expressed their deepest gratitude to all the speakers, the organizing committees, sponsors, all attendants, volunteers, etc. They also announced that the next joint workshop will be held in Guangzhou. Meanwhile, they shared their suggestions and best wishes for the future AICIT and other Â鶹´«Ã½AV Information Theory Society conferences.