novelslam Chapter 4 The Research Group's Initial Challenge
Chapter 4 The Research Group's Initial Challenge
On Monday afternoon, Zuo Cheng entered Professor Lin Zhiyuan's laboratory for the first time.
The laboratory is located on the sixth floor of the main building of the School of Telecommunications Engineering. The space of more than 200 square meters is divided into three areas: a simulation computing area, a hardware debugging area, and a small conference room. The equipment is not the latest, but it has everything you need, and the walls are covered with project progress charts and various technical parameters.
There were five people sitting in the room: four graduate students and Ma Hao.
When Zuo Cheng pushed open the door, everyone's eyes swept over him. He was all too familiar with that look—he had been scrutinized countless times in the workplace in his previous life—curiosity, sizing, and a hint of indifference.
Ma Hao leaned back in his swivel chair, holding a cup of coffee. Seeing him enter, he chuckled, "You're here? Have a seat. Professor Lin has a meeting at the department today, so she asked me to show you around."
He pointed to an empty table in the corner, on which was a thick stack of documents.
"This is the preliminary literature and raw data for this sub-direction of the research project, about 300 pages long. Please read it through first, and feel free to ask any questions." Ma Hao paused, then added casually, "By the way, this sub-direction was previously stuck on the channel modeling stage. The core difficulty is fitting the time-varying characteristics of the nonlinear channel. Our group tried two different approaches before, but neither worked. If you have any new ideas, feel free to bring them up for discussion."
After saying that, he turned back to continue his work without even giving a formal introduction.
A female graduate student wearing black-rimmed glasses smiled at Zuo Cheng and whispered, "My name is Song Yuwei, I'm a first-year graduate student. Don't be nervous, just take your time."
Zuo Cheng nodded, sat down at the table in the corner, and opened the first page of the document.
More than three hundred pages. The task Ma Hao assigned him was clearly a show of force.
At a normal reading speed, these documents would take a graduate student at least two weeks to digest. An undergraduate? Probably a month wouldn't be enough. Ma Hao was probably hoping he'd give up after three days and slink back home in disgrace.
Unfortunately, Zuo Cheng is not an ordinary undergraduate student.
He had the signal processing fundamentals and applications unlocked by the system in his mind, plus the matrix factorization ability copied from Yu Ying. He had already mastered the underlying logic of 60% of the content in these documents, and only needed to scan them to confirm the specific parameters and experimental settings.
The truly valuable part is the remaining forty percent—the section on nonlinear channel modeling.
Zuo Cheng flipped through the pages quickly, but not haphazardly. His eyes scanned each page like a scanner, capturing key information. He slowed down to examine the core formulas and experimental data carefully, while glancing over the rest.
Three hours later, he closed the last page.
Song Yuwei, who was sitting at the opposite table, looked up and saw him close the documents, and was clearly taken aback.
"You...you've finished reading?"
"I've gone through it roughly." Zuo Cheng stood up and stretched his neck. "I still need to digest the details."
Song Yuwei opened her mouth, then hesitated. Three hundred pages of documents, "roughly skimmed through" in three hours? It had taken her a full ten days to digest those materials.
Ma Hao didn't look up, but Zuo Cheng noticed that his hand, which was flipping through documents, paused for a fraction of a second.
Zuo Cheng ignored their reactions, opened the raw dataset attached to the document, and ran it on his computer. The data volume was not large, but the structure was very messy—the sampling frequency was inconsistent, some data segments had obvious noise pollution, and several timestamps did not match.
No wonder the previous two solutions failed; the data source itself was faulty.
This is like building on a crooked foundation; no matter how you build on it, it will always be tilted.
Zuo Cheng jotted down the problem in his notebook without making a fuss. He planned to clarify his thoughts first and then come up with a solution all at once—it was pointless to squeeze out the problem bit by bit; he wanted to do something so thorough that no one could argue with him.
On Wednesday afternoon, Zuo Cheng arrived at the library on time.
Yu Ying had arrived, with several pages of handwritten derivation process spread out in front of her. Her ponytail was tied neatly, and she wore a light blue knit sweater, looking as refreshing as autumn weather.
"You're here." She pushed the derivation paper towards me. "I've been following your homogeneous decomposition approach for the past couple of days, but I'm stuck on the boundary conditions—when the matrix dimension exceeds 512, the convergence speed of the decomposition drops sharply. I suspect there's a problem with the regularization term, but I haven't been able to find the root cause."
Zuo Cheng pulled the paper over and looked at it for a minute.
This is indeed a tricky problem. The convergence of high-dimensional matrices is a classic challenge in numerical analysis, and many doctoral dissertations focus on this issue.
But now he has a complete knowledge system of matrix factorization in his mind, plus the homogeneous factorization method he derived himself. Using the two sets of tools together gives him a different perspective.
He picked up a pen and drew a set of auxiliary matrices on the paper: "You see, the problem isn't with the regularization term, but with the initial conditions of the decomposition. The standard method defaults to using the identity matrix as the initial value, but for asymmetric high-dimensional matrices, this initial value is too far from the optimal solution, so the number of iterations increases dramatically, and the convergence speed naturally drops."
"What should we use as the initial value?"
"Use the lower-order approximate solution after dimensionality reduction." Zuo Cheng quickly wrote down a series of formulas. "First, project the matrix into the lower-dimensional space to calculate a rough solution, and then use this rough solution as the starting point for the higher-dimensional decomposition. It's like giving the algorithm a shortcut—no need to start climbing the mountain from scratch, just start from halfway up the mountain."
Yu Ying stared at the formula, her brows furrowing before gradually relaxing, and finally her eyes lit up.
"Dimensionality reduction preprocessing for initialization..." she murmured, grabbing a pen and quickly pushing it down a few steps. "If this holds true, the convergence speed could theoretically be improved by at least an order of magnitude."
"Well, we'll need to run data to verify the exact improvement."
Yu Ying put down her pen and looked at him seriously: "Zuo Cheng, to be honest, the maturity of your thinking is not what one would expect from a senior undergraduate."
"Maybe I started thinking about these things earlier." Zuo Cheng smiled and casually changed the subject.
Yu Ying didn't press the matter further, but there was something indescribable in her gaze towards Zuo Cheng. It wasn't suspicion, but rather a solemn reassessment.
The two discussed the issue in the library for nearly two hours, basically clarifying the boundary conditions. While collecting materials, Yu Ying suddenly said, "By the way, I heard you joined Professor Lin's research group?"
"Yeah, just got in."
"Then we're practically classmates," Yu Ying smiled at him. "I'm also under Professor Lin's supervision, but in a different sub-field. Good luck! If you can really crack that tough nut in Professor Lin's research project, it will be a huge help to you."
Zuo Cheng nodded. He noticed that Yu Ying's smile was natural and sincere, not the kind of polite social smile.
She's a good girl. In their previous life, they didn't even have a chance to meet. In this life, fate has put them on the same path.
It was already dark when Zuo Cheng came out of the library. He opened the system panel on the way and saw that the technology radar still had more than ten hours of cooldown left.
He currently has four credits. The "Fundamentals of Communication Systems" leaf on the technology tree is light gray, and he only needs one corresponding task to unlock it. He also has the research materials Professor Lin gave him—the difficult point of nonlinear channel modeling happens to be related to communication systems.
In other words, as long as he can solve the bottleneck in the research group, the system will most likely trigger the task reward simultaneously.
one stone two bird.
Back in his dorm, Zuo Cheng didn't start working immediately. He first compiled the data issues he'd discovered in the lab during the day into a concise analysis document—the specific locations of inconsistent sampling frequencies, the frequency range of noise pollution, and the nodes with misaligned timestamps. Based on this, he drafted a preliminary correction plan.
It was already 1 a.m. when they finished. Zhang Lei was snoring on the top bunk, Liu Wei was sleeping like a log with an eye mask on, and Chen Hao's computer screen was still lit, showing that he was running a set of data for a new client of Studio 402.
Zuo Cheng stretched his wrists and opened another document.
What he wants to write is a complete technical solution for fitting the time-varying characteristics of nonlinear channels.
The research group had previously tried two approaches: one based on a traditional statistical model, and the other using a neural network for hard fitting. Both methods had their drawbacks—statistical models were poor at adapting to time-varying characteristics, while neural networks required massive amounts of training data, and the research group's dataset was simply not large enough.
Zuo Cheng's approach was different. He planned to combine adaptive filtering methods and matrix factorization from signal processing to build a lightweight real-time tracking model—not aiming to calculate an exact solution in one step, but rather allowing the model to dynamically adjust according to the channel state, correcting the result of the previous step with the latest observation data at each step.
The core of this idea is precisely based on the knowledge he already possessed regarding the two blades.
He wrote until 3 a.m., completing the framework of the plan and the derivation of key formulas.
On Thursday morning, Zuo Cheng walked into the lab and placed two documents on Ma Hao's desk.
"Senior Brother Ma, this is my analysis of the original data issues, along with a preliminary plan for channel modeling. Please take a look."
Ma Hao raised an eyebrow, put down what he was holding, and opened the first document.
His expression changed when he saw the third page.
Zuo Cheng's annotations pinpointed the exact number and deviation value of each sampling point. This level of precision isn't something you can achieve by simply flipping through some materials—it requires a true understanding of the data structure and collection methods.
Then he opened the second document.
The laboratory remained quiet for a long time.
Song Yuwei secretly glanced at Ma Hao and noticed that his fingers were white from gripping the edge of the document.
"This plan..." Ma Hao began, his voice a few octaves deeper than usual, "The idea is interesting, but the theoretical derivation isn't complete enough; there are a few skipped steps. You need to fill in the gaps."
He pushed the document back and said nothing more.
But Zuo Cheng noticed a detail—when Ma Hao said "the idea is interesting," his gaze shifted slightly.
This isn't about looking down on someone.
This means they've figured out the trick, but they don't want to admit it.
Zuo Cheng took back the document and nodded: "Okay, I'll go back and finish it."
He turned and walked towards his seat, and a line of text silently appeared on the screen in his mind:
[Host detected breaking through a key technological bottleneck in the research project; new task being generated—]
[Task Name: Breaking the Channel Barrier]
[Task Description: Complete the derivation and simulation verification of a nonlinear time-varying channel modeling scheme within 7 days.]
[Mission Reward: Unlock the "Communication System Basics" blade, +5 points]
Five points added.
Zuo Cheng suppressed the smile on his lips, sat down, and opened his notebook.
Seven days is enough.
What Ma Hao didn't know was that the plan he had just pushed back for "further revisions" was already 80% complete.
The remaining 20% was simply not written down by Zuo Cheng on purpose.