I'm a scientist, why can't I have faith?

In the answer area, Lance is checking the final architecture of the model.

"That's about enough. We can make further changes if needed."

Then he checked the data processing progress and found that the results should be available in another half hour for training.

The questions were released at 10 a.m., and it's been exactly 10 hours since then.

Physically, he undoubtedly expended the most energy, not only completing the automated processing link of multimodal data based on clinical priors, but also finishing up Wei Linfeng's time-series fusion network architecture.

Although Wei Linfeng built the main framework, he worked very hard at patching and mending it.

While waiting for the data processing to finish, Lan Ce took a quick stroll to check on the progress of the other teams.

Excellent, definitely in the top tier, the strategy worked perfectly.

This is an international competition, so of course not all teams will act like Sun Yi and his team, rushing headlong into the answering area without caring about anything else.

Many teams, like Wang Dong and his group, carefully read the tournament handbook before making decisions and taking action.

Yuzu Natsuki and Kensuke Sugiyama from the University of Tokyo are among the fastest progressing students.

As expected of the dark horses who emerged from the ACM finals in the past two years, even when they gave the main computer to their teammates, the two of them worked as efficiently as anyone else, even when using laptops.

Unlike Wang Dong and his team, Natsuki Yuzu and Sugiyama Kensuke did not choose the pipeline rotation method. Instead, they designed a new algorithm for cluster quotas.

In this way, all their devices connected to the rack-mounted server array in the closed computer room of the competition venue can use the offline computing power cluster at the same time, schedule different computing power resources in parallel, without interfering with each other, and connect with zero waiting time.

However, this choice leads to a more complex algorithm design than Wang Dong's team, and it puts a great strain on the abilities of everyone on the team.

If everyone involved is a top expert, then even if the algorithm design is more complex, the efficiency of the team working together is in no way inferior to that of Wang Dong and his team.

However, if even one person lags behind, the overall progress of the project will slow down, making it undoubtedly a double-edged sword.

However, judging from their current situation, there is no possibility that they will become a burden.

Among them, Sugiyama Kensuke left a particularly deep impression on Lan Ce, not only because he was very thin and looked like a skeleton soldier, but also because he was a two-finger sparring master.

It's unbelievable to see someone coding with two fingers in the finals of an international competition. It's probably hard to find another person like that.

However, Sugiyama Kensuke's two-finger speed was in no way inferior to the ten-finger operation of an ordinary person, and Lan Ce couldn't help but admire him.

Just when Lan Ce thought this would be their biggest competitor, things took an unexpected turn.

Kensuke Sugiyama closed his notebook and called to his teammates as they walked out of the answer area.

Out of curiosity, Lan Ce stepped forward to talk to him.

"Are you only missing model training to complete the algorithm?"

"No, it's just time to rest. Don't you guys get any sleep?"

Indeed, it's already past nine o'clock at night. For people with healthy sleep schedules, it's time to take a shower and go to bed.

Upon hearing Sugiyama Kensuke's question, Lan Ce felt somewhat embarrassed and replied guiltily:

"We people of Great Xia are used to going to bed later."

The group of contestants from Japan didn't think much of it and left the answering area, stopping their countdown.

I thought I might still have a chance to bump into him, but now it seems there's absolutely nothing to worry about.

In fact, for Lan Ce, or rather for most of the Xia players present, the matter of sleeping at night was not even a consideration.

If you choose to rest, it means you are voluntarily allowing your opponent to surpass you.

Moreover, staying up all night is something these seasoned veterans are very familiar with.

Lan Ce's personal record is three consecutive nights of working through the night; otherwise, why do you think he was able to publish a paper at CVPR as the first author, even though he was an undergraduate?

This unique culture of the new generation in Daxia will probably be something the Japanese players will experience firsthand by tomorrow morning.

Lan Ce deliberately avoided Sun Yi's group during his stroll, after all, there was a plague god there, and the new students still had to be respectful.

Moreover, even from a distance, Lan Ce felt that his code might have problems.

Back at my seat, the data processing was almost complete.

"It's time to change the personnel."

He finished up the last bit and then left the answer area to find Wang Dong and the others.

At this moment, Wang Dong was doing the final finishing touches. After explaining the structure of his model to him, Wei Linfeng went back to his room to rest.

Wang Dong and Su Yuran were left to design and build a full-chain, multi-dimensional, and evidence-based interpretable system based on the algorithm.

At this point, they had completed the initial design and were finalizing their ideas. After that, all that remained was the tedious task of building the system.

The small library was noticeably more crowded than during the day.

It's not that everyone in Daxia likes to study at night, like during finals week; rather, more and more teams are realizing that if they don't work on the interpretability module in tandem, they're very likely to be unable to submit their algorithms before the deadline.

"How's it going? How far have we progressed?"

Wang Dong tidied the messy manuscripts on the table and said:

"Go through all of this again from the beginning, and all that's left is the code."

Lan Ce nodded in approval after listening.

"Then you go train the model, junior. I'll help you organize the data; it should be processed by now."

Su Yuran was stunned for two seconds, then her eyes became clear.

"Okay, I'll go now."

Finally, I don't have to worry about these thick books anymore!

She ran excitedly toward the answer area.

As the nominal captain of the team, Su Yuran is often seen as someone with connections, but she is not just a pretty face.

At least in terms of deep learning and model training, she is a true genius.

This is a fact that even Lan Ce and Wei Linfeng have to admit. Su Yuran can always do things that they cannot do efficiently and concisely, and the training and learning of the model often exceeds expectations.

Faced with this problem's algorithm, Su Yuran naturally did not choose the conventional approach. Instead, she designed a joint loss function of Focal Loss + Dice Loss for the extremely imbalanced dataset with a positive rate of 1.8%, and added a clinical rule constraint penalty term to amplify the loss weight of positive samples, thereby reducing the false negative rate from the root and avoiding the overfitting problem.

After completing the function design, she finally called the data that Lance had already processed in the offline computing cluster, and then just had to wait for the model to finish learning and training on its own.