Rebirth: Starting from Lighting Up the Tech Tree

Chapter 196 Signal Decoding

After the filtering module had been running stably for a week, Shen Yiming imported the first batch of training data into the decoding model.

The volunteer, Xiao He, was a 26-year-old administrative assistant at the research institute and was in good health. The experimental procedure was simple: he wore an EEG cap with embedded electrodes, imagined different actions according to the on-screen prompts, the system collected his neural signals, the decoding model attempted to recognize his imagined content, and the recognition results were displayed on the screen in real time.

The initial accuracy rate was 62%.

Shen Yiming was dissatisfied with the number, but not surprised. This was the starting point for the general-purpose model; further improvements would depend on data-driven model development. He adjusted the network structure, increased the number of training epochs, and ran it again. The accuracy rose to 78%. After two more epochs, it stalled at around 87%, unable to advance further.

The problem lies in individual differences. Each person's neural signal characteristics are different; even when imagining clenching a fist, the brainwave patterns of different volunteers show significant differences. The general model learns the average characteristics of all people, not the characteristics of a specific individual.

Shen Yiming had Tang Ning increase the amount of individual labeled data and collected data from three new volunteers before retraining. The accuracy climbed to 87.3, but remained within that range.

Zuo Cheng sat beside him, watching Shen Yiming adjust the parameters again and again, without saying a word. He was observing the structure of the problem.

The ceiling of a general-purpose model is essentially a dilemma of transfer learning: the signal distribution differences between different individuals are too large, and simply increasing the amount of data is unlikely to overcome this challenge. A different approach is needed, allowing the model to continuously adapt to the current user's signal characteristics during the inference phase, rather than using a fixed weight to handle everyone.

Zuo Cheng opened the system panel.

The current score is 662. He accessed the reinforcement learning module in the AI ​​framework and fused it with the EEG decoding task for analysis. The idea is to allow the decoding model to receive real-time feedback signals. Whenever the volunteer confirms or denies the recognition result, the model adjusts its parameters based on this feedback, continuously converging to the current user's signal characteristics.

System notification: Fusion is feasible, consuming 5 points, generating an adaptive reinforcement learning decoding framework, supporting online individualized calibration, and the individualized accuracy is expected to be improved to over 92%.

Zuo Cheng confirmed. Points dropped from 662 to 657.

He shared his ideas with Shen Yiming.

After hearing this, Shen Yiming's eyes lit up, and he said, "Online learning, with the model updating as it runs?"

Zuo Cheng agreed. The traditional method involves training the model offline before use, requiring data collection and retraining for each new user, which is time-consuming and costly. By using a reinforcement learning framework, the model receives feedback after each recognition, continuously updating its weights and adapting to the current user much faster.

Shen Yiming asked, "Where does the feedback signal come from?"

Zuo Cheng explained that there are two sources. One is explicit feedback, where volunteers actively confirm or correct the recognition results. The other is implicit feedback, which infers the accuracy of the previous recognition by monitoring subsequent neural signals. The first method is simpler, while the second is more accurate. We should implement the first method first, and add the second method once it's proven successful.

Shen Yiming spent three days building the framework and started testing on the fourth day.

A dozen or so people were watching the test, including Professor Zheng, Chen Minghui, and Tang Ning, with several doctoral students standing in the back. Everyone understood the significance of today's event: whether the accuracy rate could break through 90% would determine whether this system was qualified for clinical use.

Xiao He sat back down in the testing chair and put on the EEG cap. A prompt appeared on the screen: imagine raising your right hand upwards.

The model gave the recognition result: raise your right hand. Xiao He nodded, indicating that it was correct.

The prompt has been changed to imagine moving to the left.

Ten times in a row, nine were correct and one was wrong.

Shen Yiming made a note of it and then switched to the second volunteer. The second volunteer's initial accuracy rate was only 71%, but after 20 rounds of calibration, the accuracy rate climbed to 89%.

Zuo Cheng asked, "How many more rounds are needed to reach ninety?"

Shen Yiming said that at the current convergence rate, it should be able to reach the target after another thirty to fifty rounds.

They kept running.

The numbers on the screen gradually increased: 89.1, 89.7, 90.2.

Shen Yiming looked up and said, "Passed."

The number stopped at 90.2, continued to rise, climbed to 91.3, and then stabilized.

No one spoke in the lab; everyone stared at the number. Shen Yiming let out a long sigh, pressing his chair back so hard it creaked.

Professor Zheng stepped forward, looked at the number on the screen, and said, "Ninety-one point three. What does this number mean?"

Tang Ning said this means that our system has reached a world-leading level in the task of decoding motion imagination. The highest publicly reported record is currently 89.5% of that of top international laboratories, which we have surpassed.

Professor Zheng paused for a moment, then asked, "Is this ready for clinical use?"

Zuo Cheng said that the technical indicators are sufficient. The next step is to prepare the clinical trial application materials and go through the ethics approval process.

Professor Zheng nodded and said, "I'll coordinate with Huaxia University Hospital. I'm familiar with their ethics committee; the approval process usually takes about three months."

Zuo Cheng asked, "Can we recruit the first batch of volunteers in three months?"

Professor Zheng said that after approval, it will be necessary to recruit qualified volunteers. For the first batch, I suggest recruiting patients with spinal cord injuries leading to motor dysfunction, whose condition has lasted for more than a year and who have not responded to conventional rehabilitation methods. These patients have the most urgent needs, and the clinical evidence is the most convincing.

Chen Minghui said from the side, "We have made the necessary preparations in terms of hardware. The NX-30 implantation procedure has been standardized, and the surgical coordination process has been confirmed with Professor Zheng multiple times."

Zuo Cheng said, "Then let's move forward with the clinical trial application." Professor Zheng, you get it started here, Han Lu will follow up on the approval process.

Professor Zheng said, "Okay."

As the crowd began to disperse, Shen Yiming organized the test data into a document, while Tang Ning helped him check the format.

Zuo Cheng stood next to the testing chair, looking at the display screen next to him. The number on the screen was still frozen at 91.3.

Tang Ning stood behind him and said softly, "President Zuo, we really did it."

Zuo Cheng said, "This is just the beginning."

A few months later, this system will be running on real patients. At that time, the numbers on the screen will represent more than just test data; they will represent a specific person, a pair of hands eager to move again. Looking at those numbers, he knew that every step from now on would be more difficult than this one. But he also knew that nothing was more important than this step.