Cambricon Stocks Plunge After Sudden Exit of Huawei Veteran CTO; US-Banned Deep Glint Goes Public; China-US AI Collaboration Rises Fivefold Since 2010

Weekly China AI News on March 20

News of the Week

Chinese A.I. Chipmaker Cambricon Loses CTO; Stock Falls

Cambricon, a leading A.I. chip developer in China and once a key supplier of Huawei’s smartphone chipsets, announced that its CTO Jun Liang had left the company on February 10 over “a company disagreement.” Shares of the Shanghai-listed company have dropped more than 20% since the news release.

Jun Liang, who joined Cambricon in 2017 as CTO and deputy general manager, was the mastermind leading the development of Cambricon’s flagship 7-nm A.I. training chip MLU290, which is expected to challenge Nvidia’s A100. Before that, Liang served as a senior tech expert at Huawei’s chip unit Hisilicon.

“Mr. Liang Jun's resignation will have a certain impact on the company's R&D management. The company has established a complete R&D system, built a professional R&D team, and reserved a rich portfolio of patented technologies. Mr. Liang Jun's departure will not affect the company's technological innovation, nor will it have a significant adverse impact on the company's overall R&D strength,” the company said in the release.

Cambricon has lost 75% of its market value since its public debut in 2020, to only RMB26.5 billion yuan ($4.2 billion). The total revenue of Cambricon in 2021 reached RMB720 million yuan ($113 million) with a net loss of RMB850 million ($133 million), surging by 95% year over year.

DeepGlint Makes Shanghai Market Debut Raising RMB1.8B

DeepGlint, a Beijing-based computer vision and facial recognition developer, has made its public debut on the Shanghai STAR market, raising RMB1.8 billion yuan (~$282 million) at a valuation of ~RMB7.3 billion yuan ($1.2 billion). However, the stock slid by 5 percent from its opening price on the first trading day.

Founded in 2013, DeepGlint specialized in 3D computer vision and robotic perception. The company made its first buzz when Bill Gates described the company as “very cool” on his China tour in 2014. Big-name investors like Sequoia China and Zhen Fund also bet on the startup and estimated the company would eventually reach an eye-popping valuation of $300 billion.

DeepGlint didn’t live up to the expectation, however. The company reportedly hit roadblocks in developing its flagship security and surveillance system, which was designed to track movement and human behaviors in public squares. As a result, CEO and Co-Founder Bofei He was dismissed in 2017. Co-founder Yong Zhao, a Google Glass veteran, took the role and revived the embattled company. DeepGlint made a comeback in 2019 after winning an RMB100 million bid to supply surveillance cameras to the Agricultural Bank of China.

According to its prospectus, DeepGlint reported its annual revenue of RMB51.96 million yuan in 2018, RMB71.21 million yuan in 2019, and RMB243 million yuan in 2020, with a net loss of RMB74.57 million yuan, RMB418 billion yuan, and RMB78.2 million yuan, respectively.

Last year, DeepGlint was added to a U.S. government blacklist over its alleged involvement in human rights abuses.

China-US AI Collaboration Rises Fivefold Since 2010

The US-China AI race indicates a growing global competition between the two superpowers in the field of A.I., and the idea is gaining more traction over “the U.S. would soon lose the A.I. race to China” advocacy. However, Stanford University’s latest A.I. report tells a different story.

A.I. Index, an annual report to track, collate, distill and visualize data relating to artificial intelligence, said, “despite rising geopolitical tensions, the United States and China had the greatest number of cross-country collaborations in A.I. publications from 2010 to 2021, increasing five times since 2010.” However, the US-China collaborations in A.I. publications saw a slight drop in 2021.

The report also said China led the world in the number of A.I. journals, conferences, and repository publications in 2021 —63.2% higher than the United States with all three publication types combined. However, the U.S. is still a dominant player in leading among the major powers concerning the number of A.I. conference citations. You can read the full report here.

Papers & Projects

An Image Patch is a Wave: Quantum Inspired Vision MLP

A team of researchers from Huawei | Noah's Ark Lab, Peking University, and Sydney University proposed representing each image token as a wave function with two parts, amplitude and phase. Based on the wave-like token representation, the research team established a novel Wave-MLP architecture for vision tasks. Extensive experiments demonstrate that the proposed Wave-MLP is superior to the state-of-the-art MLP architectures on various vision tasks such as image classification, object detection, and semantic segmentation. You can read the paper on arXiv.

Scaling Up Your Kernels to 31x31: Revisiting Large Kernel Design in CNNs.

A team of researchers from Tsinghua University and Megvii revisited large kernel design in modern convolutional neural networks (CNNs). Inspired by recent advances of vision transformers (ViTs), in this paper, they demonstrate that using a few large convolutional kernels instead of a stack of small kernels could be a more powerful paradigm. They proposed RepLKNet, a pure CNN architecture whose kernel size is as large as 31x31, in contrast to commonly used 3x3. RepLKNet greatly closes the performance gap between CNNs and ViTs, e.g., achieving comparable or superior results than Swin Transformer on ImageNet and a few typical downstream tasks, with lower latency. You can read the paper on arXiv and code on GitHub.

Deep learning guided optimization of human antibody against SARS-CoV-2 variants with broad neutralization

To catch up with the virus’s evolution, a team of Tsinghua University researchers, UIUC and MIT introduced a deep learning approach to redesign the complementarity-determining regions (CDRs) to target multiple virus variants and obtained an antibody that broadly neutralizes SARS-CoV-2 variants. Read the paper on PNAS.

Rising Startups

Aibee, a Beijing-based AI solution startup, has raised $100 million from China Merchants Capital and smartphone maker Xiaomi, which propels its valuation to more than $1 billion. Founded in 2017 by Baidu veteran Yuanqing Lin, the company aims to upgrade the vertical industries through A.I. total solutions with a near-term focus on the retail market.

Abrobo, a Shenzhen-based surgery robot startup, has raised RMB120 million yuan ($18.8 million) in its Series Pre-A funding round. Founded in 2020, the company specializes in vascular interventional surgical robots, which can be used for percutaneous coronary intervention (PCI), peripheral vascular intervention (PVI), and neurointerventional surgery (NVI).

Rest of the World

DeepMind introduced a variant of its large language model Gopher, namely GopherCite, which can generate answers whilst also citing sources to support its claims. The 280 billion-parameter model performed well in multiple question-answering datasets under human evaluations but DeepMind researchers also admitted evidence citation plays a limited role in providing trustworthy answers. You can read the blog for more.

John Bernal was a Tesla employee who owned Tesla’s self-driving beta system FSD. However, he was fired recently after he uploaded an FSD review video on his YouTube Channel showing Tesla hit bollards. Tesla also deprived him of access to the FSD Beta system in the vehicle he owns. You can read the full story on CNBC.

A team of researchers from MIT trained a quadrupedal robot to run much faster. Unlike other agile running controllers behind Boston Dynamics' robot, which is programming-based, they used reinforcement learning and digital simulation to make the university’s robot Mini Cheetah run at 3.9m/s or roughly 8.7mph. Watch the video below.