Rebirth: Starting from Lighting Up the Tech Tree

Chapter 110 Federation

Two weeks later, Shen Yiming knocked on Zuo Cheng's office door.

"President Zuo, the engineering of the federated learning framework is complete." Shen Yiming's tone carried a hint of barely perceptible excitement. "Ma Hao helped me optimize the model aggregation algorithm, and Fang Ze's team has successfully run the Cambricon engineering sample; all three-line parallel tests have passed."

Zuo Cheng put down the documents in his hand: "Can we proceed with the actual device now?"

"Yes, it's possible. But it can only be tested on a small number of nodes first; large-scale deployment requires further verification."

"Then let's try it today." Zuo Cheng stood up. "Call Chen Hao and Fang Ze, and go to the smart city's control center."

Half an hour later, Zuo Cheng and his group stood in front of the large screen at the Hangzhou Smart City Control Center.

The screen displays real-time data from sensors across the city: 1,500 edge gateways and 5,000 sensor nodes distributed across roads, bridges, waterways, and buildings in six districts. Data streams, like fine rays of light, continuously converge on the central platform.

"Choose a gateway in one district for testing," Zuo Cheng said. "West Lake District, 120 gateways."

Shen Yiming started working on his laptop. He distributed the federated learning training task to the edge gateways in the West Lake District. Each gateway used only local sensor data to train a small model, then encrypted the gradient information and uploaded it to the aggregation server. The aggregation server then aggregated the gradients from all the gateways and updated the global model.

"Begin," Zuo Cheng said.

Data began to flow on the screen. One hundred and twenty edge gateways simultaneously began local training, and gradient information was encrypted and uploaded after each round. The progress bar of the aggregation server slowly climbed from zero.

"First round of aggregation complete, model accuracy 78%." Shen Yiming stared at the data. "Second round begins."

Zuo Cheng stood behind and quietly opened the system panel. The technology radar had finished cooling down, and Shen Yiming was just three meters away.

scanning.

The panel displayed the results. Compared to two weeks ago, Shen Yiming had added another skill to his list: gradient sparsity compression. This was a new capability he developed himself during the engineering process, which could remove redundant data from gradient information, retaining only key parameters, thereby significantly reducing communication overhead.

Zuo Cheng looked at the technology and quickly calculated in his mind. If gradient sparsity compression were replicated, combined with Shen Yiming's existing adaptive communication scheduling, the communication efficiency of federated learning could be improved to another level.

[Copying "gradient sparsity compression" requires 8 points. Continue?]

confirm.

The points increased from 285 to 277. A new technology appeared in the list of tree leaves in Zuocheng's technology tree, and was categorized under the Internet of Things branch.

Zuo Cheng closed the panel and continued watching the test.

"The third round of aggregation is complete, with the model achieving 85% accuracy." Shen Yiming's voice trembled slightly. "The data has never left the local system, ensuring complete privacy protection. Fourth round."

"What about communication costs?" Chen Hao asked.

"It's 52 percent lower than centralized training," Shen Yiming said. "This data has exceeded my expectations. When testing on the Raspberry Pi, we could only achieve a 37 percent reduction. The real-world environment is actually better because the edge gateway's computing power is much stronger than the Raspberry Pi's, resulting in higher gradient quality during local training and thus better aggregation efficiency."

"What about the loss of accuracy?" Fang Ze asked.

"Almost none." Shen Yiming pulled up a comparison chart. "The accuracy of the model trained by federated learning is 85%, while the accuracy of the model trained by centralized learning is 87%. The difference is within the error range. Moreover, this difference is narrowing as the number of training rounds increases."

Zuo Cheng's heart skipped a beat. He had already copied gradient sparsity compression from the system, but the carrier of this technology was not Shen Yiming, but his own technology tree. He needed to "feedback" this technology to Shen Yiming, but he couldn't reveal its source.

"Yiming, have you continued to work on the gradient sparsity approach you mentioned before?" Zuo Cheng asked.

Shen Yiming paused for a moment: "I have some initial ideas, but I haven't had time to implement them yet."

"I read a paper before about adaptive threshold selection for gradient sparsity. The core idea is to dynamically adjust the sparsity rate during training, retaining more gradients in the early stages to ensure convergence, and increasing the sparsity rate later to reduce communication. You could try this approach."

Shen Yiming's eyes lit up: "Adaptive threshold selection? This is a great idea, and it can be integrated with my previous adaptive compression ratio. I'll implement it as soon as I get back."

Testing continued. After the eighth round of aggregation, the model accuracy stabilized at 93%, almost on par with the accuracy of centralized training. Communication overhead finally stabilized at a reduction of 58%.

"Test successful." Shen Yiming breathed a sigh of relief, his hands trembling slightly. "The federated learning framework has run successfully on the real device. The data never left the local machine, the accuracy is almost lossless, and communication overhead has been reduced by 58%."

Chen Hao patted him on the shoulder: "Well done, treat the team to a good meal when you get back."

Fang Ze rarely showed a slight smile, then continued to look down at the power consumption data, muttering, "Peak power consumption is 12 percent lower than estimated. Cambricon's chip is alright."

Zuo Cheng stood in front of the large screen, watching the data streams converging from the edge gateway. Federated learning, model compression, and Cambricon chips—three technology lines were converging. Technological amplification was working silently in the background, increasing the actual efficiency of all integrated solutions by twenty percent.

But he didn't mention it to anyone.

Just then, my phone rang.

Zhou Henian's name appeared on the screen.

"Zuo Cheng, there's something I need to tell you." Zhou Henian's voice was as direct as ever. "The technical review for Phase IV of Sky Dome needs to be brought forward. Blue Bay wants to include AI capabilities as a core indicator for Phase IV. How is your AI team in 402 preparing?"

Zuo Cheng gripped his phone tightly, his heart racing.

"Mr. Zhou, the AI ​​capabilities of 402 are ready to be deployed at any time."

"Okay. Come to Blue Bay next week for a technical demonstration."

After hanging up the phone, Zuo Cheng looked out the window at the sky. The Federation Learning Framework had just completed its first practical application, and Blue Bay's Sky Dome Phase IV had arrived. The timing was perfect.

However, the AI ​​requirements of Phase IV of Sky Dome are far more complex than the spectrum scheduling of Phase III. Zuocheng needs stronger technological reserves.

He glanced down at the system panel. Two hundred and seventy-seven points, four branches and thirty-nine leaves on the tech tree, and the tech radar was still cooling down.

Is that enough?

Zuo Cheng didn't know. But he did know one thing: the AI ​​in 402 was no longer an empty shell.