miHoYo exits, the Shanghai Jiao Tong University-backed investment fund takes over Lingweiyisi | 36Kr Exclusive

By Hu Xiangyun

Edited by Hai Ruojing

At the end of October, miHoYo has withdrawn from the shareholder roster of Shanghai Lingweiyisi Technology Co., Ltd. (hereinafter referred to as "Lingweiyisi"). 36Kr has learned that this brain-computer interface company founded by Lv Baoliang, a professor at Shanghai Jiao Tong University, has officially completed its restructuring and secured exclusive investment of several million yuan from Hanyuan Asset of Shanghai Jiao Tong University. The funds will be used for team expansion and research and development of the foundational electroencephalogram (EEG) model.

Lingweiyisi was founded at the end of 2021. It initially started with affective brain-computer interface, exploring the application of EEG and eye movement signals in the objective assessment of emotional disorders such as depression and anxiety, and had previously launched the multimodal emotion assessment system "Emotion X-ray Machine" and the foundational EEG model.

Shi Lichen, co-founder and general manager of the company, told 36Kr that in the future, the company will expand from affective brain-computer interface technology to foundational EEG models and general neural decoding platforms, adopting a strategy that combines underlying technology R&D and application exploration in parallel. Shi Lichen holds a doctorate degree from the School of Computer Science of Shanghai Jiao Tong University, and studied under Professor Lv Baoliang. He once served as technical director of Huolala, chief algorithm engineer of Didi, and senior engineer of Baidu, with 6 years of entrepreneurial experience.

According to Shi Lichen, this expansion "is an extension from specific applications to underlying common technologies, rather than switching to a completely new track". For traditional EEG decoding algorithms, the feature expression and task decision-making systems are trained simultaneously, which can only use task-related labeled data. The high cost of data labeling, limited data scale and diversity lead to limited decoding capabilities. In addition, the algorithm is highly bound to equipment and tasks. When switching to a different device or scenario, data collection and model training often need to be restarted from scratch. "For example, when switching from a sleep monitoring task to an emotion recognition task, or replacing an EEG acquisition device, data collection, labeling and model training all need to start all over again."

From the perspective of the Lingweiyisi team, the foundational EEG model may provide a solution. Its core idea is to split EEG decoding into two stages. The first stage uses self-supervised learning to learn general neural representations in EEG signals on a large scale on unlabeled data; the second stage conducts post-training with a small amount of labeled data for specific tasks, so that the model can form decision-making capabilities on the basis of existing representations. This two-stage "pre-training-post-training" framework draws on the idea of decoupled training of feature expression and decision-making systems, and adapts it to the special signal modality of EEG.

"Pre-training uses unlabeled data, which has much lower collection cost. Moreover, the larger the scale and the more diverse the covered devices and crowds, the richer and more generalized the learned neural representations will be," Shi Lichen explained, "In the post-training stage, we only need to learn the decision boundary, which greatly reduces the demand for labeled data. This means that when facing new users, new devices or new applications, the adaptation cost and sample dependence of the model are expected to be reduced."

This logic can be analogized to the mobile communication industry. Before the emergence of smartphones, in the feature phone era, hardware, operating systems and application software were highly bound, and switching devices meant developing applications from scratch. However, smartphones realized the "adaptation" between hardware and applications through unified operating system interfaces, which gave rise to the explosion of the mobile Internet.

The Lingweiyisi team believes that the brain-computer interface industry is still in the "feature phone stage" to some extent at present. "Therefore, we hope to separate the neural decoding capability from specific scenarios and turn it into a reusable underlying capability, so that different devices can be connected at low cost, and application developers do not have to solve EEG algorithm problems from zero."

In terms of technology, the team's existing technical reserves mainly come from the previous research on foundational EEG models by Lv Baoliang's team.

First of all, the paper related to the LaBraM model, which is used to solve the problem of general representation of EEG signals, has been published at ICLR2024, a top conference on machine learning. "LaBraM has achieved certain results in various tasks such as EEG anomaly detection and emotion recognition, which initially verified the path of building a reusable foundational EEG model through large-scale pre-training." Recently, the team has been continuously promoting large-scale research on foundational EEG models, carrying out pre-training based on nearly 80,000 hours of multi-source EEG data, to further explore the cross-task generalization capability under larger data and model scales.

The second work, NeuroLM, further explores unified multi-task decoding. It converts continuous EEG signals into sequence representations similar to language tokens, and combines with large language models to support multiple EEG tasks through instruction training.

"For example, traditional models may need to develop two separate sets of independent algorithms for emotion recognition and EEG anomaly classification respectively. When we were developing NeuroLM, we tried to use a unified model to complete corresponding decoding through different task instructions. This study used about 25,000 hours of EEG data, and the maximum model scale reached 1.7 billion parameters. Its significance lies in exploring whether EEG decoding can evolve from multiple independent dedicated models to a unified multi-task model," said Shi Lichen.

In addition, Lv Baoliang's team launched an intelligent system named CerebraGloss this year, which mainly focuses on the professional understanding and generative analysis of clinical electroencephalogram. In the past, when clinical doctors read electroencephalograms, they needed to observe complex waveform features combined with experience, while traditional AI usually can only identify simple problems, such as whether a certain segment of EEG is abnormal. "CerebraGloss tries to use the vision-language large model to analyze electroencephalograms, which can not only identify relevant waveforms, but also describe and answer questions in natural language, and support multi-round interaction."

According to Shi Lichen, these three works are intended to "explore the capability boundary of large EEG models from different directions". At the same time, Lingweiyisi is promoting the iteration of the training paradigm for the next-generation foundational EEG model, and plans to introduce multi-modal data such as external stimuli, perception, cognition, and behavior that are time-aligned with EEG.

In addition to the exploration at the model level, the Lingweiyisi team is also continuously accumulating its own data set. At present, the team has accumulated more than 100,000 hours of EEG data, including both large-scale data from public channels and nearly 10,000 cases of self-collected data.

In terms of commercialization, Shi Lichen said that the ultimate application of brain-computer interface has not yet been finalized, and Lingweiyisi plans to carry out technical platform cooperation and application exploration simultaneously. On the one hand, it will continue to improve the decoding capability of the foundational EEG model and its adaptation capability across devices and applications; on the other hand, it will quickly verify the technical value through scenarios such as medical treatment, AI interaction, and embodied intelligence. "Later, we will launch an available intelligent development platform for brain-computer interface, providing neural decoding services to partners in the form of API, SDK or private deployment."

Investor's View:

Chen Zhihong, Deputy Investment Director of Hanyuan Asset, said: Lingweiyisi's strategic upgrade from affective brain-computer interface to foundational EEG model accurately grasps the evolution of the brain-computer interface industry from "dedicated tools" to "general intelligent platforms". The team has solid research foundation in the field of brain-computer interface, and has many years of practical experience in entrepreneurship and serving as technical leaders in top-tier tech companies, which injects industrialization acceleration into the company, and this is the variable we value extremely. Relying on the interdisciplinary ecology of Shanghai Jiao Tong University in brain science, AI and clinical medicine, Hanyuan Asset will provide all-round empowerment for Lingweiyisi in algorithm iteration, clinical verification and industrial resource docking, and promote the accelerated implementation of its general EEG decoding platform.