Information and Engineering Breakthroughs: Network Coding, the Shannon Award, and the Journey from the Multimedia Laboratory to SenseTime

CUHK's engineering research boasts two world-class high-water marks of strikingly different character: one is network coding, pioneered by Raymond Yeung — a new field in information theory with a CUHK scholar as co-founder, which earned information theory's highest honour, the Claude E. Shannon Award, in 2022; the other is Tang Xiao'ou's Multimedia Laboratory, which made landmark contributions at the inflection point of the deep learning wave in 2014 and directly incubated the AI unicorn SenseTime. The former "rewrote a field's body of knowledge"; the latter "turned laboratory results into industry". This article is a fact-based archival entry in the reference zone (Section 04, Research), carrying no credibility badge, with every claim individually linked to academic, official, or secondary sources. The scholars involved are public figures, named truthfully according to public records (Tang Xiao'ou passed away in 2023).

---

# Part One · Network Coding: A New Information Theory Field Founded at CUHK

# 1. What Is Network Coding: A Counter-Intuitive Insight

In the conventional view of network communication, nodes in a network do exactly one thing: store and forward — receive a data packet and pass it on unchanged, like a post office sorting letters. Network coding proposes a counter-intuitive insight: an intermediate node may "encode" (mathematically combine) the packets it receives before forwarding them, and the receiving end then "decodes" to recover the originals — which, in certain network topologies, can significantly improve transmission efficiency.

According to the IEEE Information Theory Society's Shannon Award announcement^※ and English Wikipedia^※, network coding theory was jointly proposed by Raymond W. Yeung and collaborators in the late 1990s, and "fundamentally changed how people understand network communications". Its foundational paper "Network Information Flow"^※ was published in 2000 (co-authored by Yeung with R. Ahlswede, Ning Cai, and Shuo-Yen Robert Li). The theory's canonical demonstration is the butterfly network — in this topology, an intermediate node that only forwards creates a bottleneck; if encoding is permitted, the needs of multiple receivers can be satisfied simultaneously, approaching the network capacity limit.

A paradigm shift: the significance of network coding lies not in inventing a specific protocol, but in redefining "what nodes in a network can do" — elevating them from passive "porters" to active "computers". This is an important extension of information theory along the "network" dimension since Shannon.

# 2. The Arithmetic of the Butterfly Network: Where the Counter-Intuition Lies

An abstract phrase like "encoding by combination" is easy to gloss over, so it is worth working through the classic butterfly network. Suppose a source wants to multicast two messages, a and b, to two receivers; each edge in the network can carry only 1 bit per unit time, and there is a bottleneck edge in the middle shared by both receivers. Under "store and forward", this bottleneck edge can only choose to carry either a or b at any given moment, so one of the two receivers always ends up missing one of a and b — throughput is choked at the bottleneck.

Network coding's approach is to have the intermediate node at the bottleneck avoid the either-or choice and instead send a⊕b (the bitwise XOR of the two bits). One receiver, which has already received a by another route, uses a⊕(a⊕b) to recover b; the other, having received b, uses b⊕(a⊕b) to recover a. One XOR, and both receivers are satisfied simultaneously — precisely what the foundational paper Network Information Flow^※ proved: in multicast scenarios, allowing encoding can approach the theoretical capacity bound set by the max-flow min-cut theorem, which conventional routing generally cannot attain. This "⊕" may look like primary-school arithmetic, but it is the sharpest cut of the entire theory: it shows that information is neither water nor parcels, that it need not be conserved and shuttled around, but can be superimposed and then re-split.

The difference between "carrying" and "computing" can be put in one sentence: routing treats each packet as an indivisible physical object, network coding treats it as a number that supports algebraic operations. The former's ceiling is how many physical objects the min-cut can pass; the latter's ceiling is how much information the min-cut can pass — and in multicast, the latter is often higher.

# 3. The Founder and CUHK's Role

According to English Wikipedia^※ and CUHK's official press release^※, Yeung is CUHK's Choh-Ming Li Professor of Information Engineering, co-director of the University's Institute of Network Coding, and a co-founder of network coding as a field. This means network coding is not a piece of research that "a CUHK scholar happened to participate in", but a brand-new field with CUHK as one of its key seedbeds.

Yeung's academic trajectory makes the provenance of this "CUHK mainstream line" clear. According to English Wikipedia^※, he earned his bachelor's, master's, and doctoral degrees in electrical engineering from Cornell University in the United States in 1984, 1985, and 1988 respectively, under the supervision of information theorist Toby Berger; after his doctorate he worked in the Performance Analysis Department at AT&T Bell Laboratories from 1988 to 1991, then returned to Hong Kong in 1991 to join CUHK, where he has remained for over three decades. In other words, the seeding of network coding as a discipline connects on one side to the orthodox information theory lineage of Shannon and Bell Labs, and on the other side takes root on the hilltop campus of Hong Kong. In 2010, he took the lead in establishing the Institute of Network Coding at CUHK, funded by the UGC's Areas of Excellence (AoE) scheme (see Research Output and Spin-off Companies), extending a highly theoretical direction into practical scenarios such as data storage and network communications.

# 4. From Existence to Linear Network Coding: Three Papers Build a Discipline

Network coding was not settled by a single paper in one stroke; rather, several successive works progressively consolidated "what is feasible" into "what is usable".

• 2000 · Existence: Network Information Flow^※ (Ahlswede, Cai, Li, Yeung) proved that in multicast networks, coding can achieve the max-flow min-cut bound — it answered the question "is it possible?";
• 2003 · Linearity suffices: Linear Network Coding^※ (Li, Yeung, Cai; published in IEEE Trans. Inf. Theory) further proved that complex nonlinear coding is unnecessary — linear operations alone (treating data blocks as vectors over a finite field, with nodes performing linear transforms) suffice to achieve the multicast capacity bound — it answered "how simple a method is enough", pulling the theory from "some encoding exists" to "an engineering-realizable encoding";
• Randomisation · Distributed operation: subsequently, random linear network coding (RLNC) proposed by Ho, Médard and others allowed each node to choose its coding coefficients independently at random, achieving successful decoding with high probability without global coordination, enabling network coding to run in real distributed, dynamic networks. MIT's COPE protocol ("XOR in the air") measured substantial throughput gains in wireless mesh networks, taking the blackboard's a⊕b into real routers.

These three papers trace a clear descending path: from "coding is useful" (2000) to "linear suffices" (2003) to "random works and can be distributed" — difficulty decreasing step by step, usability rising step by step. CUHK scholars are named authors on the first two rungs, which is precisely what "co-founding" means in practice.

# 5. Theoretical Depth: From the Zhang–Yeung Inequality to BATS Codes

Beyond network coding itself, the team has made several further contributions to both the foundations and applications of information theory. According to English Wikipedia^※:

• The Zhang–Yeung inequality: around 1998, Yeung and Zhen Zhang discovered this inequality together, establishing the existence of the first "unconstrained non-Shannon-type information inequality" — an important breakthrough in the foundations of information theory. For decades beforehand, it had been assumed that the set of inequalities Shannon gave in 1948 exhausted the constraint relationships among entropies; the Zhang–Yeung inequality proved that when four or more random variables are involved, new constraints exist that Shannon's framework fails to capture. According to English Wikipedia^※, this revealed the incompleteness of the constraints on entropy functions and opened an entire subsequent line of research;
• Machine-proving entropy inequalities: alongside this discovery, Yeung also pioneered the use of computers to automatically prove information inequalities, developing tools including ITIP (Information Theoretic Inequality Prover, 1996) and AITIP (2020), converting the question of "which information inequalities hold" from a hand-derived problem into one solvable by software as linear programming;
• BATS codes (BATched Sparse codes): he invented BATS codes, which improve transmission rates over networks that suffer packet loss — a representative result of network coding moving toward engineering applications, systematically treated in his monograph BATS Codes: Theory and Practice (2017), co-authored with Shenghao Yang. Teams such as n-hop technologies, incubated at CUHK, are applying these codes to real products for multi-hop wireless transmission.

"Foundations × Applications" two-wheel drive: the Zhang–Yeung inequality belongs to the most abstract fundamental theory (the mathematical constraints on information measures), while BATS codes confront the most practical engineering problem (transmission over lossy networks). The same team working at both the theoretical ceiling and the applied ground floor is the hallmark of top-tier information theory research.

# 6. Off the Blackboard: Where Network Coding Is Used

Network coding is often mistaken for pure theory, but its applications are in fact wide-ranging. According to CUHK's official introduction to the Hamming Medal^※, the IEEE awards committee determined that Yeung's work "has permanently changed the way we view network communications, with impact on error-correcting codes, information security, wireless communications, and data storage" — this official characterisation neatly points to the four directions into which network coding has permeated:

This table also explains why an ostensibly abstract branch of information theory has been repeatedly written into the engineering literature of wireless, storage, and security over two decades: what it changes is not some protocol but the underlying assumption about "what nodes can do with data" — move the foundation, and many layers above must be redrawn.

# 7. Textbooks, a MOOC, and "Training a Generation"

A top scholar's influence lies half in papers, half in teaching materials. Yeung's information theory textbook is reported to be adopted by over 100 universities worldwide; according to English Wikipedia^※, his publications include A First Course in Information Theory (2002), Network Coding Theory (2005, co-authored with Shuo-Yen Robert Li and others), and the comprehensive Information Theory and Network Coding (2008) — the last of which compiles Shannon's classical information theory and the newborn network coding into a single textbook, itself a marker that "network coding has become a formal chapter of information theory". From 2014 onward he put the course on Coursera, reportedly attracting over 60,000 learners in total.

A plain criterion for whether a field has "established itself" is whether it has a widely recognised textbook and a course everyone takes. Network coding has both — and both come from the hand of its co-founder. The person who opened the field also wrote its introductory primer.

# 8. Recognition: The 2022 Claude E. Shannon Award and Three Medals

According to the IEEE Information Theory Society^※ and CUHK's official press release^※, Yeung received the 2022 Claude E. Shannon Award, honouring his "sustained and profound contributions to information theory", and delivered the Shannon Lecture at ISIT in Finland that year. Named after Claude Shannon, the founder of information theory, the Shannon Award is the highest honour in the field — roughly the equivalent of a "lifetime achievement supreme award" for the discipline.

The Shannon Award is no isolated coronation; it is the culmination of a string of medals. According to UGC Awards^※ and CUHK's official pages^※:

The 2016 Sumner Award was shared with Ning Cai, a collaborator who later taught at Xidian University — the award trail faithfully preserves the authorial order of those foundational co-authored papers. From Fellow to Shannon Award, the keyword recurs across all four honours: network coding. A CUHK scholar reaching this summit marks the University's world-class standing in information theory as a foundational discipline.

---

# Part Two · From the Multimedia Laboratory to SenseTime: An Industrialisation Legend of Deep Learning

# 9. Asia's Only AI Laboratory in the Global Top Ten

At the heart of CUHK's world-class standing in artificial intelligence (especially computer vision) is the Multimedia Laboratory (MMLab) in the Faculty of Engineering. According to English Wikipedia, "Tang Xiao'ou"^※, the MMLab was founded by Tang Xiao'ou, a professor in CUHK's Department of Information Engineering, and was at the time one of the world's top ten artificial intelligence laboratories and the only one located in Asia — likewise the only Asian representative on Nvidia's 2016 list of the world's ten leading AI laboratories.

According to the MMLab website^※, the laboratory was founded by Tang in July 2001. Alongside him, its core research and teaching strength included Wang Xiaogang and Chen Change Loy, both of whom later achieved considerable renown in their own right. Its world-class status is no empty claim: the lab won the CVPR 2009 Best Paper Award, which by its own account was the first — and only — time in thirty years that the top conference awarded best paper to an Asian research institution. It subsequently won best-paper awards at NeurIPS, CVPR, ACM MM, COLT and others, and in 2014 became Hong Kong's first Nvidia CUDA Research Center. Its graduates have gone on to MIT, Stanford, Cambridge, Columbia, Carnegie Mellon and other leading institutions, earning the laboratory a reputation in the industry as a "Whampoa Military Academy" (黄埔军校) of computer vision.

In a landscape where AI research is heavily concentrated in North America, a Hong Kong university laboratory ranked "top ten globally, the only one in Asia" is in itself powerful testimony to CUHK's engineering research strength. This laboratory was the seedbed of the SenseTime saga that followed.

# 10. From GaussianFace to DeepID: Twice Crossing the Human Baseline

The MMLab's most iconic achievement is the DeepID face recognition algorithm of 2014. According to English Wikipedia, "SenseTime"^※ and the DeepID academic paper (CUHK EE)^※, in 2014 the team released the face recognition algorithm DeepID, described as the world's first algorithm to surpass human accuracy in face recognition. Building on deep learning, the breakthrough stands as one of the landmark moments in the history of computer vision.

To grasp the weight of "surpassing humans", one needs an accepted yardstick. The academic community has long used the LFW (Labeled Faces in the Wild) dataset of "in-the-wild" faces as the benchmark for face verification, on which human accuracy is about 97.53% — that is the "human-level line". GaussianFace^※, also from the CUHK team and published at AAAI 2015, where it won an Outstanding Student Paper Award, reported an accuracy of 98.52%, crossing that human-level line for the first time; the DeepID series that followed pushed accuracy even higher. According to the Overview of Research Achievements, citing CUHK in Focus, DeepID reached 99.15% on comparable benchmarks, surpassing Facebook's DeepFace from the same period and becoming a landmark result in deep-learning face recognition. That same year, the laboratory's DeepID-Net also placed second in the object-detection track of the ImageNet Large Scale Visual Recognition Challenge (ILSVRC) 2014, sometimes called the "Visual Olympics".

The significance of "surpassing humans": face recognition had long been considered a task "machines could hardly match humans at". The CUHK team first pushed machine accuracy past the human-level line with GaussianFace, then pulled further ahead with the DeepID series — not a single lucky crossing, but two successive generations of methods standing at the very forefront of this global technological inflection point.

# 11. SenseTime: From Laboratory to Industry

DeepID's technical breakthrough swiftly translated into industrial force. According to English Wikipedia, "SenseTime"^※, in October 2014, SenseTime was co-founded by Tang Xiao'ou and computer scientist Xu Li, among others, as a research-driven spin-off incubated from the academic research of CUHK's Multimedia Laboratory. It subsequently grew into one of the most representative enterprises in China's computer vision sector, listing on the Hong Kong Stock Exchange in late 2021. According to the source, the CUHK–SenseTime Joint Lab has published or presented over 400 computer vision papers in top-tier academic journals and conferences — a number second only to Microsoft. In other words, the "CUHK–SenseTime" pairing of academia and industry ranks among the very top globally in computer vision research output.

It is worth noting that this incubation line is not SenseTime alone. Under CUHK's knowledge transfer system (ORKTS), the same community of computer vision and information engineering faculty and students has successively spawned additional unicorn or near-unicorn enterprises including SmartMore and Insighta (see Research Output and Spin-off Companies) — the Multimedia Laboratory and the Department of Information Engineering have in effect become an "upstream source of talent and technology" for AI entrepreneurship in the Greater Bay Area.

The "academia-industry integration" paradigm: the SenseTime story is a complete loop of "laboratory breakthrough → spin-off enterprise → joint lab feeding back into academia" — the laboratory produces technology and talent, the enterprise provides application scenarios and resources, and the joint lab converts applications back into new academic results.

# 12. A Necessary Boundary Note

As a listed company, SenseTime's operations, valuation, and commercial controversies are corporate matters that change with the market and with time; this museum does not track or judge them. This article records only two traceable facts: the "academic achievements of CUHK's Multimedia Laboratory" and "the incubation lineage of SenseTime as an academic spin-off". Conflating a company's commercial performance with a laboratory's academic contributions would be inaccurate and is not this museum's forte. Readers interested in SenseTime's commercial affairs should consult its financial reports and regulatory disclosures as first-hand sources.

---

# Part Three · Two High-Water Marks Converging

Placing network coding and DeepID/SenseTime back into CUHK's research lineage, they join optical fibre and NIPT in forming the University's map of "original breakthroughs with far-reaching impact", each with its own emphasis:

If optical fibre, network coding, and NIPT are more about "rewriting a field's body of knowledge", then DeepID/SenseTime more vividly demonstrates the path of "directly incubating laboratory results into industry", echoing the Greater Bay Area's logic of "Hong Kong basic research + mainland China industrial translation". A university's most far-reaching industrial influence sometimes lies less in how many companies it has founded than in one of its laboratories happening to stand at the inflection point of a technological era — in the 2014 deep learning wave, CUHK's Multimedia Laboratory was in the thick of it, and contributed outstandingly.

# One Department, Two Kinds of World-Class

Viewed side by side, the two high-water marks harbour an easily overlooked coincidence: they come from the same department. Both the Institute of Network Coding and the Multimedia Laboratory sit under the Department of Information Engineering in CUHK's Faculty of Engineering — one department, with one arm reaching into the most abstract information theory (entropy, inequalities, the existence of codes) and the other into the hottest deep learning (faces, detection, super-resolution). This is not an administrative accident but two expressions of the same disciplinary genes: information theory asks "where are the limits of information", machine vision asks "how to approach those limits from pixels", and both share a common language of "representation, compression, redundancy".

The two leaders' models of contribution, however, form a mirror image. Yeung took the slow road of "theory first, applications later": the Zhang–Yeung inequality of 1998 and the foundational papers of 2000 and 2003 would wait more than a decade, until RLNC and BATS codes matured, before revealing themselves in wireless and storage — the reward being a new discipline with CUHK as one of its seedbeds, and a Shannon Award. Tang took the fast road of "breakthrough means industry": DeepID crossed the human baseline in 2014, and in October of that same year SenseTime was established, converting the accuracy figures in a paper directly into a company valuation. Between slow and fast there is no hierarchy — they represent two legitimate ways of doing top-tier research: one pushes knowledge toward ever further boundaries, the other swiftly delivers knowledge at the boundary into reality.

The ideal picture of a research university is perhaps precisely this side-by-side view: it can accommodate one person spending twenty years proving an inequality, and another converting a paper into a company within a year. The former determines how deep a university can go; the latter, how broadly it can reach — and in the same department, both happen at once.

Related reading: Overview of Research Achievements, Research Output and Spin-off Companies, State Key Laboratories, Breakthroughs in Life Sciences and Medicine.

---