ICCV 2021 Best Paper; First Delivery Robot Accident in China; How NLP May Improve mRNA COVID-19 Vaccine

China’s AI news in the week of October 17, 2021

Swin Transformer Wins ICCV 2021 Best Paper

A premier conference in computer vision, the 2021 International Conference on Computer Vision took place virtually last week from October 11 to 17. The top honor, the ICCV 2021 Best Paper or Marr Prize, was awarded to Swin Transformer: Hierarchical Vision Transformer using Shifted Windows from the University of Science and Technology of China, Xian Jiaotong University, and Microsoft Research Asia.

Swin Transformer is a variant of Vision Transformers (ViT) that serves as a backbone of computer vision (as convolution does for vision and Transformer does for NLP). Researchers made two improvements upon existing ViT architectures:

• They proposed a hierarchical architecture like convolutional neural networks to better represent visual signals at various scales and enjoy linear computational complexity concerning image size;

• They introduced Shifted Windows, which brings greater efficiency by limiting self-attention computation to non-overlapping local windows while allowing for cross-window connection.

Swin Transformer achieves state-of-the-art performance on the recognition tasks of image classification, object detection, and semantic segmentation:

• 87.3 top-1 accuracy on ImageNet-1K;

• 58.7 box AP and 51.1 mask AP on COCO test- dev;

• 53.5 mIoU on ADE20K val.

The ICCV committee accepted 1612 papers from 6152 submissions, an acceptance rate of 26%. 43.2% of the paper submissions came from China, followed by the U.S. (23.6%). Transfer learning, few-shot learning, and unsupervised learning are the top three subject areas of accepted papers.

Meituan Is at Fault in Delivery Robot Accident

On October 9, a low-speed autonomous delivery robot (“robot”) developed by Chinese food delivery giant Meituan collided with another human-driving car (“car”) in Beijing, Chinese media reported, making it the first publicly-known delivery robot accident in China.

The robot was driving on a side road - a non-motor vehicle lane - and about to cross a T-junction straight when its right side bumped into the left front of the car that was turning right at the moment. Meituan explained that the robot did an emergency brake while the car didn’t stop immediately. This, however, conflicted with a driving record provided by the car owner.

Meituan was held full responsibility for the accident and damages over the violation of driving on a non-motor vehicle lane, according to a traffic accident report. Local regulations rule delivery robots are considered neither motor vehicles nor non-motor vehicles, a regulatory grey zone that puts Meituan’s self-driving tech in a dilemma.

The robot, dubbed Modai 20, can drive with a maximum speed of ~28 mph and carry up to 330 lbs, said the company. Equipped with three LiDARs, 19 cameras, two milliwave radars, and nine ultrasonic radars, the robot can navigate within a distance of between 5 centimeters to 150 meters.

Meituan, JD.com, and Chinese startup Neolix received permits to test delivery robots on public roads starting this May amid an increasing adoption of driverless tech to ease logistics pressure.

AI Algorithm Designs Stable RNA for Covid-19 Vaccines

mRNA has emerged as a pioneering vaccine technology amid the raging COVID-19 pandemic. Still, its instability remains an hurdle that leads to stringent cold-chain requirements and two-dose injection to reinforce immune responses. This week, Baidu announced that their AI algorithm could help improve the stability of mRNA COVID-19 vaccines by designing a much better RNA.

Baidu’s preclinical studies with StemiRNA, a Shanghai-based mRNA biotechnology company, showed that their algorithm LinearDesign can produce better mRNA vaccine sequences for SARS-CoV-2 Spike protein in the stability, protein expression, and immunogenicity.

The LinearDesign algorithm was inspired by a natural language processing technique called incremental parsing. Parsing is a critical process for computers to understand languages by dividing a sentence into multiple words or phrases - noun, verb, determiner - that make semantic sense. However, the length of time to understand a sentence scaled cubically with sentence length so that if you doubled the size of a sentence, it took eight times longer to parse it.

In a recent editorial piece on IEEE Spectrum, Baidu scientists explained that their innovative incremental parser can scan a sentence from left to right, output many possibilities, and choose the most likely one. This method significantly reduces the time required for parsing linearly to the length of the sentence.

So how is that related to COVID? Like a sentence, RNA consists of a long string of nucleotides. The coronavirus genome, which is a type of RNA, contains some 30,000 nucleotides. The number of possible RNA sequences encoding the spike protein could reach about 2.4 x 10 to the power of 632. Using traditional methods, the time of predicting RNA structure grows exponentially along with the length of an RNA. Baidu’s approach achieves the same improvement in RNA analysis as incremental parsing does on NLP.

You can learn more from the paper: LinearDesign: Efficient Algorithms for Optimized mRNA Sequence Design.

Investment News:

• Hua Kong Tsingjiao, a Beijing-based information science company, founded by Tsinghua University, has raised RMB 500 million yuan (~$77 million) in its Series B financing round. Established in 2018, the company focuses on researching, developing, and deploying privacy-preserving computing by incorporating modern cryptography, game theory, and AI. For example, it has built a secure multi-party computing platform for China Everbright Bank.

• SJ Semiconductor, a leading MEOL (Middle-End of Line) foundry, has raised $300 million in Series C funding. Founded in 2014, the company is a joint venture between Semiconductor Manufacturing International Corp. (SMIC) and Jiangsu Changjiang Electronics Technology Co., Ltd (JCET). It aims to offer first-class MEOL manufacturing and testing services and develop the advanced 3D IC business.

• Whale, a digital marketing company, has raised $50 million in its Series B funding round. Founded in 2017, the company has built a team of ex-Facebookers, ex-McKinseys, and ex-IBMs with strong backgrounds in AI, IoT, data, consulting, and commercial digital transformation. It provides AI solutions and marketing cloud to 200 clients such as Nio, Watsons, and Uniqlo.