Scientists from Peking University dive into brain-computer interface, securing nearly 100 million yuan in seed round financing

Text | Sun Xiaowen

Interview / Editing | Hai Ruojing

Exclusive information obtained by "Hidden Currents Waves" reveals that the invasive brain-computer interface company "Xinsheng Vision" recently completed a seed-round financing of nearly 100 million RMB. This round of financing was led by Matrix Partners China, with follow-on investments from Xinglian Capital, Yanyuan Venture Capital, and Shuimu Venture Capital.

Currently, invasive brain-computer interfaces have been applied in medical scenarios such as treating paralysis and brain-controlled peripherals, validating the safety and effectiveness of long-term implantation. Meanwhile, the accelerated evolution of AI Agent and embodied intelligence technologies has also magnified the market's expectations for brain-computer interfaces: to rewrite the future of "human-machine interaction", achieve human-machine integration, and enhance humanity. Elon Musk and his Neuralink are the vanguards of this narrative.

The Xinsheng Vision team brings together several "Young Chief Scientists" in the field of brain-computer interfaces under the national "Brain Project": the founder, Wang Qian, is from the School of Psychological and Cognitive Sciences and the IDG/McGovern Institute for Brain Research at Peking University. The co-founder, Li Yuanning, is from ShanghaiTech University. He is a doctor in artificial intelligence and neuroscience from Carnegie Mellon University and has conducted postdoctoral research for many years in the Edward Chang Laboratory (a top laboratory in brain-computer interfaces) at the University of California, San Francisco. The co-founding team also includes experts from Peking University in the fields of embodied intelligence, brain-computer chips, and brain-inspired chips.

This young team of scientists has received investment simultaneously from first-tier technology funds and institutions affiliated with Peking University and Tsinghua University right from the start. Their chosen direction is clear: the integration of vision, language, and embodiment; the goal is to develop a "neural graphics card" brain-computer interface system with ten thousand channels and one-piece implantation.

Li Yuanning, the co-founder of Xinsheng Vision, believes that vision is the most important perceptual input for humans, while language is the core output for human expression. In brain-computer interface research, the key questions are how to efficiently write visual information into the human brain and how to read out intentions and semantics. This corresponds to the functional reconstruction of blind and aphasic patients and is closer to the core issue of human-machine interaction.

Why choose visual reconstruction as the first step, and what level can be achieved at present? What exactly does the "neural graphics card" in brain-computer interfaces mean? How can vision, language, and embodied intelligence form the bridge for the next generation of human-machine interaction? The following is our edited conversation with Li Yuanning:

01 Scientists Enter the Arena

Hidden Currents: How did the co-founders of Xinsheng Vision come together to decide to establish the company?

Li Yuanning: This actually stems from a shared judgment on the technological endgame: the essence of human interaction with the world lies in "perception" on one end and "expression" on the other.

Wang Qian and I have been researching the visual and language cortices of the human brain for a long time. More than a decade ago, during my doctoral studies at CMU, I conducted invasive research on the neural encoding and decoding of the human visual cortex at the University of Pittsburgh Medical Center and published several studies on text and memory hallucinations induced by electrical stimulation. Later, I also participated in leading the earliest invasive Chinese brain-computer interface research in China.

Meanwhile, Professor Wang Qian's team was the first in China to achieve single-neuron recording in the visual cortex based on clinical patients and published a research paper in 2020 on "The mapping mechanism of color visual hallucinations induced by electrical stimulation of the human visual cortex".

When we gathered the underlying algorithms and experimental data accumulated over a decade and wanted to cooperate to truly promote their clinical transformation, we found that we urgently needed a hardware base that could be precisely regulated to achieve a two-way closed-loop of information.

For this reason, we joined forces with the top team at Peking University that specializes in brain-computer hardware chips, who have rare industrial-grade tape-out and large-scale mass production experience.

Another co-founder of Xinsheng Vision, Professor Zhu Yixin, also from Peking University. As one of the few experts in embodied intelligence in China who has been invited to give a report at the NVIDIA GTC Global Conference, he was also impressed by the vision of "human-machine integration" and joined the team.

By then, the team had finally gathered experts across the entire chain, from neural mechanism analysis, underlying high-throughput chips, cutting-edge clinical medicine, to embodied AI interconnection. This is also our technological barrier from silicon-based computing power to carbon-silicon integration.

Hidden Currents: There are already several brain-computer interface companies in China valued at over 10 billion RMB. Isn't it a bit late to enter the market now? Can the data and capabilities accumulated by previous companies in motor function reconstruction be easily generalized to indications such as vision and language?

Li Yuanning: In terms of the time of company establishment, we are not among the early ones. However, most previous companies have focused on motor function reconstruction, while we have chosen the directions of vision, language, and embodiment, where everyone is just starting out, and our accumulation may be even deeper.

The data flywheel accumulated in motor function reconstruction is difficult to transfer to other cognitive functions.

The dimensionality of motor control is relatively low, but vision and language involve extremely high-dimensional non-linear encoding. In this new arena, Xinsheng Vision's accumulation lies not only in algorithms but also in the fact that our founding team has been on the front line of clinical practice for many years, accumulating a large amount of intracranial multi-modal data, including a database of electrical stimulation of the visual cortex of over a hundred patients. This means we don't need to start clinical trials from scratch.

02 The Possibility of Ten Thousand Channels

Hidden Currents: Several invasive brain-computer interface companies currently favored by the market mostly started their businesses with the advantage of "flexible electrodes". What is Xinsheng Vision's solution?

Li Yuanning: Flexible electrodes are the physical interface to enter the brain, which is very important. However, a brain-computer interface is backed by a whole system. What truly determines the upper limit of brain-computer interaction and information throughput are the chips and neural encoding and decoding algorithms beneath the electrodes - that's why we proposed the concept of "neural graphics card".

Some hardware currently used in clinical practice has low controllability and is not designed for efficient two-way interaction. Just as graphics cards have supported the explosion of AI, the "neural graphics card" system that Xinsheng Vision is building aims to solve the "encoding-decoding-reconstruction" closed-loop of extremely high-bandwidth data streams in the brain. It has the potential to become the core infrastructure supporting brain-computer integration and human-machine symbiosis.

Hidden Currents: There is also a view in the industry that a brain-computer interface system does not necessarily require extremely high throughput. When Elon Musk's Neuralink started, it implanted a brain-computer system with 1024 channels, but in reality, only a few hundred channels were effective.

Li Yuanning: Different tasks have different throughput requirements. The dimensionality of neural signals needed for motor control is relatively low. For example, to control a mouse precisely on a two-dimensional plane, less than ten degrees of freedom of neural signals for arm movement are sufficient.

In essence, all current language decoding work also decodes the fine movements of articulatory organs: only a dozen degrees of freedom of neural signals are needed to well characterize the movements of articulatory organs such as the lips, tongue, and pharynx and decode and generate speech.

However, true visual reconstruction and natural semantic decoding are projects on a different scale. If you want to write a clear image pixel by pixel into a blind person's mind or directly decode "a complex and abstract thought flashing through the mind", each step corresponds to the concurrent processing of thousands or even tens of thousands of dimensions of information. At this time, the necessity of high-throughput electrodes and high-throughput chips becomes apparent. Currently, Musk's Neuralink can achieve 3000 channels, and we are working towards ten thousand channels.

Hidden Currents: So you think it is necessary to develop high-throughput brain-computer chips. What are the difficulties in making brain-computer chips?

Li Yuanning: Developing brain-computer chips is like "dancing in shackles". To implant such chips into the brain, they need to be small in size, low in power consumption, capable of high-throughput real-time concurrency, and support two-way interaction.

This problem is not insoluble. Compared with GPUs, the current manufacturing process of brain-computer chips is not as advanced. However, without large-scale production, the profit margin will be very thin. Chip research and development is very capital-intensive, and a single small-scale tape-out may cost millions.

The field of brain-computer interfaces is just in its infancy. The amount of resources invested determines how many problems can be solved. From this perspective, compared with fields such as chips and embodied intelligence, the amount of funds received in the brain-computer interface field is still relatively small.

Hidden Currents: What is the current progress of your brain-computer chip solution?

Li Yuanning: We are one of the few teams in China that have successfully taped out a 28nm brain-computer chip. The current chip has 256 channels, and we should produce a higher-throughput, two-way interactive chip this year.

03 The Future of Visual and Language Reconstruction

Hidden Currents: Why has "visual reconstruction" become an important direction that everyone wants to tackle now? What is your team's accumulation in this area?

Li Yuanning: From the perspective of clinical demand, the demand for visual reconstruction is large enough, and there is currently no good alternative solution. The three most common types of disabilities in China are physical, visual, and hearing disabilities. For physical disabilities, there are solutions such as prosthetics, and for hearing impairments, cochlear implants can partially solve the problem. However, in the field of visual impairment, for completely blind patients, the implanted brain-computer interface is almost the only solution for restoring vision.

From the perspective of the research path, the first-step verification route is clear enough. The primary visual cortex has a retinotopy mechanism, which means that stimulating the corresponding position in the cortex can produce color vision and light perception.

Professor Wang Qian started relevant research as early as 2020, asking subjects to describe where in their field of vision a picture appeared after receiving cortical electrical stimulation. For example, after dividing the field of vision into several parts, patients could perceive red on the left and blue on the right. Thanks to the advantages of domestic clinical resources, more than 100 cases of experimental data have been accumulated, partially solving the problem of simple bottom-up input in the visual cortex.

In terms of visual coding algorithms for complex scenarios, there is a biennial international challenge in the field of neural coding called Algonauts. The organizers provide a public neural big data set and require participants to submit algorithms to predict the brain's neural response activities as accurately as possible based on the given stimulus content. Our team won the global championship in 2021 and 2023. However, we did not participate in 2025, and the champion that year was Meta AI.

Hidden Currents: Previously, some brain-computer interface experts said that the difficulty of visual reconstruction is extremely high. Why is there a huge challenge in "restoring vision to the blind"?

Li Yuanning: Because there are still two key problems that have not been solved.

Firstly, it is the mechanism by which the brain complements the environment in dynamic vision. When we look at the world, our brain actually only receives visual information from a very small area at each moment: the area with high resolution in our retina is very small. The brain relies on the high-frequency saccades of the eyeballs, combined with the internal world model, to fill in and piece together a complete and continuous visual scene.

This mechanism also makes the reconstruction of brain-computer interfaces extremely difficult. Since the primary visual cortex is not a simple feedforward network, there are a large number of lateral inhibitions and feedback connections across brain regions. The activation of one neuron can laterally affect surrounding neurons, which means that activating one area may affect another area, increasing the difficulty of visual coding reconstruction. If we simply and rudely activate ten thousand electrodes simultaneously as ten thousand independent pixel points, we will probably get a blurry image.

Another difficulty is how to use neural plasticity to enable the brain to learn to accept such input and generate perception. When we input electrical signals encoded by algorithms into a blind person's brain, for the brain, this is a new type of "language" it has never seen before.

Blind people cannot immediately "see" the world after connecting to the device. Instead, they need to go through an active adaptation process similar to "reinforcement learning". In essence, this is to use neural plasticity to enable human-machine co-evolution and ultimately help the brain "grow" a new perceptual channel.

Hidden Currents: What is the significance of the reconstruction and research of visual and language functions in brain-computer interfaces for the future of "human-machine interaction"?

Li Yuanning: Language and vision are the most core input and output for our interaction with the world. Vision provides about 80% of our perceptual input and builds a complex "world model" deep in the brain; language is responsible for decoding these consciousnesses and outputting them externally.

However, language itself is actually a low-bandwidth, lossy compression constrained by the physical limitations of human motor organs. The ultimate goal of Xinsheng Vision is to break this bondage and achieve high-throughput human-machine interaction, building a direct bridge between the high-dimensional semantic space of the human brain and large models and embodied AI.

In the future, the interaction of thoughts may no longer need to be reduced to explicit language or actions. The human brain may be able to achieve direct semantic connection with AI. This may not only be the endgame of human-machine interaction but also an evolutionary leap for carbon-based life and silicon-based intelligence to cross physical boundaries and move towards symbiotic integration.