At ShanghaiTech, encouraging undergraduates to step into labs and dive into research isn’t just a slogan—it’s a tradition that’s been around since day one. Recently, Wang Yuhan, a Class of 2023 undergraduate from the School of Information Science and Technology (SIST) and now a PhD student at the Chinese University of Hong Kong, made waves with two research achievements from his undergraduate years. Both works have been accepted by top international journals in communication and information theory, IEEE Journal on Selected Areas in Communications and IEEE Transactions on Information Theory, respectively. Not only do they shine in theory, but they also hold promise for real-world applications in intelligent communication system and distributed computing. Guided by Associate Professor Wu Youlong from SIST, the two papers both have Wang Yuhan as the first author, with ShanghaiTech as the primary affiliation.
“Task-oriented lossy compression”: one squeeze, multiple wins
The first study tackles a classic challenge in communication systems: how do we compress data efficiently when bandwidth is tight? Traditional methods, like the “information bottleneck” (IB), are specialists—they squeeze out redundant data to serve a single goal, such as classification. But when we need to juggle multiple tasks—like reconstructing an image, generating content, and identifying objects—they start to stumble.
A lossy compression model that comprehensively considers distortion, perception, and classification tasks
In this study, the team came up with a clever fix. They introduced two new tools: “rate-distortion-classification” (RDC) and “rate-perception-classification” (RPC), turning the conventional theory into a multitasking expert. Picture this: you’re compressing a photo and need it to stay clear enough to reconstruct (distortion), while still keeping the key details for a machine to recognize what’s in it (classification). The proposed approach smartly balances these goals. The team discovered that an excessive focus on achieving crystal-clear images (low distortion) may compromise classification accuracy, and under conditions of limited bandwidth, this trade-off becomes increasingly challenging. Using a Gaussian source model, they pinned down how noise affects the system performance and built a deep learning-based framework to handle multiple tasks at once, proving it works with real data.
RDC and RPC functions under a Gaussian source
This tech is like an intelligent regulator for compression. Take autonomous driving: cars need to compress data under tight bandwidth, ensuring maps reconstruct clearly while preserving critical details for road detection. That’s where “task-oriented lossy compression” shines!
“Coded distributed computing”: cracking the puzzle of messy networks
The second study dives in on distributed computing—think a team of computers splitting a big job. The trick is cutting down on “chitchat” between them. Existing coded distributed computing (CDC) schemes can pack data cleverly to save communication load, but they rely on neatly pre-set rules for data and task assignments. In real life, like edge computing, devices are a chaotic bunch—some hold this data, others want that result, and there’s no central planner to line it all up.
The team tackled this mess with two new solutions: “one-shot coded transmission” (OSCT) and “few-shot coded transmission” (FSCT). They used a “deficit ratio” to figure out what each device has and needs, then split and packed the data smartly, boosting efficiency with multicasting (sending one message to multiple recipients). OSCT operates as an independent decoder—each data packet can be processed individually, making it well-suited for scenarios requiring rapid response. FSCT, conversely, functions as an integrated optimizer, processing multiple packets collectively to maximize resource efficiency, rendering it ideal for setups prioritizing peak performance. Through theoretical analysis, they nailed down the conditions for these schemes to hit the sweet spot between computation and communication load.
This is like a “fuel-saver mode” for distributed systems. In edge computing—say, a smart home where devices work together but don’t synchronize perfectly—this approach slashes communication costs.
From classroom to cutting edge: ShanghaiTech’s research journey
Though these studies focus on different fields, they exhibit a “theory-to-practice” orientation. In the compression work, the team started with information theory, built a unified framework, and made it real with deep learning. In the computing study, they modeled chaotic systems and crafted scalable algorithms. Wang completed the core theory and experiments during his undergraduate years, demonstrating exceptional research proficiency.
The success is attributable to ShanghaiTech’s hands-on methodology. The SIST often assigns project-based “big homework,” teaming undergraduates with graduate students to tackle cutting-edge problems from an early stage. Courses integrate theoretical foundations with contemporary research, fostering a reciprocal relationship between education and innovation.