Chinese Cities Race to Win Self-Driving Game; Tencent Launches New AI Hearing Aids; Huawei Proposes Challenging Autonomous Driving Dataset
Weekly China AI News from Sep 19 to Sep 25
Dear readers, sorry for skipping the last week’s issue due to my personal matters. This week, I talked about how Chinese cities are scrambling to lift regulatory restrictions on autonomous vehicles amid a state-wide push to commercialize self-driving tech. Plus, on the World Day of The Deaf, Tencent offered a new affordable AI hearing aid with better sound quality and denoising. I also presented three terrific papers on AI, including a super hard autonomous driving dataset from Huawei that fails SOTA object detectors.
News of the Week
Chinese Cities Embrace Self-Driving Cars with Open Arms
What’s new: Three major Chinese cities announced new regulatory approvals last week that will allow semi-self-driving cars and robotaxis to operate in cities, joining China’s ongoing push to make road autonomy a reality.
- Wuxi, a city near Shanghai in eastern China, expands the test and operation area for autonomous vehicles (AVs) to the whole city, claiming to be the first Chinese city to do so. However, a human driver is still mandated to sit behind the steering wheel. WeRide.ai, a Guangzhou-based AV startup, quickly echoed the news by moving its eastern China headquarter to Wuxi and promising a fleet of robotaxis and delivery cars to run on the ground.
- Suzhou, a city west of Shanghai known for its classical gardens, has been home to GM-backed Momenta and ByteDance-backed QCraft, both of which have launched pilot robotaxi programs. Pony.ai has joined the competition, announcing a joint partnership with Geely-backed Caocao Mobility and Geely’s research institute to deploy a robotaxi pilot service in Suzhou by the year-end.
- Guangzhou, one of the four Tier-1 cities in China, gave Chinese EV maker XPeng the green light to launch its pilot program of City Navigation Guided Pilot (NGP), which is an FSD-like Advanced Driver Assistance System (ADAS) feature that can navigate through complex urban roads with human supervision required. City NGP will be first made available on the premium version of XPeng’s sedan P5.
Why it matters: Testing and trial operations of robotaxis or semi-autonomous vehicles will become a new normal after China issued a new guideline this August that encourages the adoption of autonomous vehicles and buses, in easily controlled scenarios. Recode China AI also reported that AV is one of the ten AI applications China committed to supporting.
Tencent Improves Hearing Aid with AI
What’s new: On September 26, the last Sunday of September on which World Day of The Deaf is celebrated, Tencent introduced a new AI-powered hearing aid that aims to help people — mainly the aged with hearing loss — to hear better.
Why it matters: China is home to 120 million people aged over 65 with some degrees of hearing loss, according to a local survey. Over 63 million people need the assistance of hearing aids, but only 6.5% of them are well-equipped with devices. The barriers that prevent them from purchasing hearing aids include unaffordable prices and quality issues.
Chinese Internet giants are also assuming more social responsibilities and promoting social values against the backdrop of the tech crackdown in the country.
How AI algorithms could help: Low latency, low power consumption, and high sound quality characterize a favorable hearing aid. Named “Zhiting 挚听” Tencent’s new hearing aid lowers its sound processing latency to 10 milliseconds and reduces its power consumption. As a result, the hearing aid has improved its speech clarity and comprehensibility by 85% in complex scenarios.
AI algorithm is also developed to eliminate the disturbing whistling or squeaking noise, which occurs when the sound loops back around and goes into the hearing aid’s microphone.
Denoising — especially adaptive denoising — is another critical feature of hearing aids. In addition to a denoising algorithm for environmental sound, the “Zhiting 挚听” can also recognize and reduce hundreds of common noises in everyday life scenes, such as the sound of washing, closing doors, chopping, and cups falling on the ground, etc. Thus users won’t be frightened.
scBERT as a large-scale pretrained deep language model for cell type annotation of single-cell RNA-seq data
Researchers from Tencent AI Lab, SJTU, and the University of Texas at Arlington developed a pretrained deep neural network-based model, single-cell bidirectional encoder representations from transformers (scBERT) to overcome challenges in annotating cell types based on single-cell RNA-seq. Extensive and rigorous benchmark studies validated the superior performance of scBERT on cell type annotation, novel cell type discovery, robustness to batch effects, and model interpretability. The paper has been published in Nature Machine Intelligence.
SegNeXt: Rethinking Convolutional Attention Design for Semantic Segmentation
Researchers from Tsinghua University and Nankai University presented SegNeXt, a simple convolutional network architecture for semantic segmentation. The paper says convolutional attention is a more efficient and effective way to encode contextual information than the self-attention mechanism in transformers. SegNeXt outperforms EfficientNet-L2 w/ NAS-FPN and achieves 90.6% mIoU on the Pascal VOC 2012 test leaderboard using only 1/10 parameters of it. The code is available on GitHub.
CODA: A Real-World Road Corner Case Dataset for Object Detection in Autonomous Driving
Researchers from Huawei introduced a challenging dataset named CODA that consists of 1500 carefully selected real-world driving scenes, each containing four object-level corner cases (on average), spanning more than 30 object categories. On CODA, the performance of standard object detectors trained on large-scale autonomous driving datasets significantly drops to no more than 12.8% in mAR. The paper has been accepted at ECCV 2022.
Evomotion, a Shenzhen-based full-stack AI solution provider, has raised $20 million in its Series B financing round. Founded in 2015, the company is specialized in developing edge-based neural processing units (NPU) for IoT. The built-in NPU of its first AIoT chip, EvoSense ONE provides 3.0 TOPs of computing and reduces power by about 30% compared with mainstream processors.
Tpson, an Hangzhou-based AI company, has raised several hundreds of millions of RMB in its Series C financing round. Founded in 2015, the company offers AI software-as-a-service to monitor electricity data and detect fire hazards.