US to Sanction Chinese Video Surveillance Giant; Alibaba Open Sources Federated Learning Framework; Drone Swarms Navigate Through Heavy Forest

Weekly China AI News from May 2 to May 8

News of the Week

U.S. To Sanction Chinese Video Surveillance Giant Hikvision

What’s news: The U.S. government is mulling imposing an unprecedented sanction against Hikvision, the world’s largest video surveillance company, over alleged human rights abuse, Financial Times reported. The company’s stock plunged by 20 percent since then.

More details: The U.S. Department of Treasury will implement the sanctions, the report says. The Treasury Department enforces the well-known Specially Designated Nationals And Blocked Persons List (SDN), in which assets of blacklisted entities and groups are blocked and U.S. persons are generally prohibited from dealing with them. It’s unknown what sanctions will be employed.

Who is Hikvision? Founded in 2001, Hikvision is a Chinese state-owned manufacturer and supplier of video surveillance equipment for civilian and military purposes. The Hangzhou-based companies supply its cameras to over 1,000 cities from more than 180 countries.

The company has already been added to an “entity list” in 2020, which forbids it from using US technology without approval, and another “Chinese military-industrial complex companies” list, which prohibits Americans from investing in the businesses.

Company’s response: Hikvision released a statement saying “the potential action by the US government, as reported, remains to be verified. We believe any such sanction should be based on credible evidence and due process…We look forward to being treated fairly and without bias.”

What’s the consequence: “An SDN designation would vault Hikvision past Huawei to become the most-sanctioned Chinese tech company,” wrote Jon Bateman, a fellow in the Cyber Policy Initiative of the Technology and International Affairs Program at the Carnegie Endowment for International Peace. “The most severe sanctions would freeze all of Hikvision’s assets and create civil and potential criminal penalties for anyone globally who sends Hikvision money or property…In other words, U.S. sanctions could stop Hikvision from selling anything to or buying anything from all countries friendly with (or at least afraid of) the United States.”

Alibaba Open Sources Federated Learning Platform

What’s new: Alibaba this week released an open-sourced federated learning platform named FederatedScope, which is said to provide convenient usage and flexible customization for various federated learning tasks in both academia and industry.

What is federated learning? Federated learning is a machine learning technique that trains an algorithm across multiple decentralized edge devices or servers holding local data samples, without exchanging them. Introduced by Google in 2017, federated learning gains growing traction due to the rising concerns over data privacy and ownership. Google also open-sourced TensorFlow Federated in 2019 for machine learning on decentralized data.

How it works: FederatedScope aims to address the challenges brought by heterogeneity. Uneven distribution of data across different computation devices, plus different data privacy strengths and learning goals, will weigh on the effectiveness of federated learning models. That’s why FederatedScope characterizes an event-driven architecture, in which “the participants wait for certain events (e.g., model’s parameters are broadcast to the clients) to trigger corresponding handlers (e.g., training models based on the local data). Hence, users can express the behaviors of servers and clients independently from respective perspectives rather than sequentially from a global perspective, i.e., the case in a procedural programming paradigm. As a result, it becomes easier for users to implement FL algorithms, and the implementations are finer modularized.”

The platform also features other comprehensive functionalities like auto-tuning, privacy protection, and personalization.

Code: You can visit GitHub and read the paper FederatedScope: A Flexible Federated Learning Platform for Heterogeneity.

Drone Swarms Navigate Through Heavy Forest

What’s new: Researchers from Zhejiang University developed a swarm of miniature but fully autonomous drones that can navigate through heavy forests without the help of GPS. The paper, Swarm of micro flying robots in the wild, has been published in Science Robotics.

What’s the challenge? Drones, or aerial robots, have been deeply immersed in our day-to-day lives for video recording and communications purposes, as well as industrial usages like outdoor inspection and surveying. Speaking of aerial swarm systems, the first thing that came to my mind is the drone light show recently performed at Tesla’s Texas Gigafactory grand opening. However, aerial swarm systems, particularly a swarm of drones exploring unforeseen areas with obstacles, remained unsolved.

Potential applications: Such a task is found critical in real-world scenarios like providing disaster relief. For example, a swarm of drones can search and rescue trapped people in a forest or a collapsed city in earthquakes. Scientists can also leverage the capabilities of drones to conduct biodiversity studies in the Amazon Rainforest.

Solutions: The research team developed a trajectory planner for swarm drones that functions in a timely and accurate manner based on limited information from onboard sensors. “The planning problem satisfies various task requirements including flight efficiency, obstacle avoidance, and inter-robot collision avoidance, dynamical feasibility, swarm coordination, and so on, thus realizing an extensible planner. Furthermore, the proposed planner deforms trajectory shapes and adjusts time allocation synchronously based on spatial-temporal joint optimization. A high-quality trajectory thus can be obtained after exhaustively exploiting the solution space within only a few milliseconds, even in the most constrained environment. The planner is finally integrated into the developed palm-sized swarm platform with onboard perception, localization, and control.”

Papers & Projects

Aesthetic Text Logo Synthesis

In the paper Aesthetic Text Logo Synthesis via Content-aware Layout Inferring, researchers from Peking University and Tencent proposed a content-aware layout generation network that takes glyph images and their corresponding text as input and synthesizes aesthetic layouts for them automatically. They constructed a dataset named TextLogo3K, consisting of about 3,500 text logo images from Tencent’s video platforms and their pixel-level annotations. They used Conditional GAN to synthesize text logos and developed a dual-discriminator module, including a sequence discriminator and an image discriminator, to evaluate both the character placing trajectories and rendered shapes of synthesized text logos, respectively.

BOAT: Bilateral Local Attention Vision Transformer

In the paper BOAT: Bilateral Local Attention Vision Transformer Researchers from Baidu and the University of Hong Kong proposed Bilateral lOcal Attention vision Transformer (BOAT), which integrates feature-space local attention with image-space local attention to reduce the expensive computation costs of Vision Transformers. Experiments showed that by integrating BOAT with both Swin and CSWin models, the BOAT-CSWin model clearly and consis- tently outperforms existing state-of-the-art CNN models and vision Transformers.

Rising Startups

1KMXC, a high-tech R&D company for AI robot solutions, has raised hundreds of millions of dollars in its Series D funding round. Founded in 2013, the Hangzhou-based Alibaba-backed company provides robotic car wash service that covers 138 cities across China as of the end of 2021.

Orca-Tech, a manufacturer of self-driving boat, has raised RMB50 million in its Series A+ funding round. Founded in 2017, the Shanxi-based company develops autonomous watercrafts for river cleaning and tourism.