In traditional understanding, deep learning requires strong computational support, and most of the training and inference are done in cloud data centers. Different terminal devices obtain AI capabilities from the cloud via the network. However, factors such as high latency and security risks in the cloud significantly impair the AI functionality experience of terminal devices. The migration of AI from the cloud to the edge is an inevitable trend, and embedded AI is rapidly rising!
Embedded AI allows devices to perform large-scale computations locally without relying on cloud data centers for AI capabilities, enabling real-time environmental perception, human-computer interaction, and decision control without an internet connection. It has broad application prospects in smart homes, robotics, autonomous driving, security, and industrial inspection.
Embedded devices have limited computing power and are extremely sensitive to power consumption and cost. To give embedded devices AI capabilities, we must first resolve the contradiction between computing power, power consumption, and cost.
In terms of computing power, emerging AI chip companies are launching neural network processors to provide computational support for embedded devices through heterogeneous computing combined with traditional MCUs. Established semiconductor giants like NXP, STMicroelectronics, and Renesas Electronics are also actively exploring the AI direction of MCUs.
At the algorithm level, the industry has developed many mature and effective optimization methods for neural network models, such as model compression, pruning, low-precision quantization, and the design of lightweight neural network models specifically for embedded devices like SqueezeNet, MobileNet, and ShuffleNet, which can effectively reduce power consumption in embedded devices. Furthermore, many internet and semiconductor giants have launched terminal inference engines to provide one-stop solutions for embedded AI developers in terms of neural network model compression, conversion, and adaptation to different platforms.
Thanks to continuous optimization and breakthroughs in computing power and algorithm ecosystems, the development and prosperity of embedded AI applications are highly anticipated, but they also face many difficulties.
In discussions with many embedded AI developers, it has been found that embedded AI application development is an extremely complex engineering task that requires developers to consider processor selection, open-source framework choices, neural network model design, optimization, deployment, etc., and possess solid and comprehensive hardware and software knowledge along with rich practical experience. Currently, embedded AI still has a long way to go, requiring the participation of players across the entire industry chain, including chips, algorithms, and applications, to improve.
To this end, Zhidx has launched a new embedded AI series, focusing on computing power, algorithms, application development, and deployment optimization. The content covers processors, development boards, AI development platforms, IoT operating systems, the design, optimization, and deployment of lightweight neural network models, terminal inference engines, and practical application development, providing a systematic explanation of mainstream embedded AI products and development practices.
The first batch of the Embedded AI series officially launched 4 lectures, starting from March 5. Additionally, all 4 lectures will be upgraded to video live broadcasts.
We have invited four technical experts: Qiu Jianbin, head of the Rockchip Toybrick AI development platform; Zhang Xianyi, CEO of Pengfeng Technology; Xiong Puxiang, founder of RT-Thread; and Tong Zhijun, partner & CTO of Yuemian Technology, to provide in-depth explanations on AI chips, development boards, AIoT operating systems, and lightweight neural network model design. Among them:
On March 5, Qiu Jianbin, head of the Rockchip Toybrick AI development platform, will discuss how “AI development platforms can help embedded developers accelerate application commercialization”. He will provide a systematic explanation of the current status and challenges of embedded AI development, the design of the Rockchip Toybrick AI development platform, and practical applications in retail scenarios such as customer flow statistics and DMS (Driver Monitoring System).
On March 17, Zhang Xianyi, CEO of Pengfeng Technology, will present on “Using AI development boards to achieve embedded visual application development”, systematically explaining the design experience, challenges, differences among mainstream AI development boards, and practical development of dual-camera face recognition and facial panel machines.
On March 24, Xiong Puxiang, founder of RT-Thread, will discuss “The requirements and challenges of operating systems for embedded AI application development”. He will explain the challenges faced by IoT operating systems, response strategies, and future development trends based on the competitive landscape of global IoT operating systems.
On April 17, Tong Zhijun, partner & CTO of Yuemian Technology, will present on “Designing lightweight neural network models for embedded devices”. He will showcase the development history of neural network model design from “feature-driven”, “data-driven”, “precision-first” to “speed-first”, and explain how to achieve efficient deployment and operation of neural network models on embedded devices through practical cases.
The Embedded AI series will continue to be updated, so please look forward to more explanations.
Target Audience
1. Algorithm engineers and researchers in machine learning, deep learning, CV, etc.
2. Embedded engineers, embedded software developers, and researchers
3. IC designers
4. Students in related fields such as IoT, communication engineering, electronics, automation, and computer science
Joining Path
For each session, we will set up a main speaker group and invite the speaker to join. By joining the main speaker group, you can not only listen to the live broadcast for free and receive course materials in advance, but also directly meet and communicate with the speaker. Of course, you can also connect with more technical experts.
Friends who want to join the main speaker group can scan the QR code at the bottom of the poster to add assistant Bei Bei (ID: zhidxioe) to make an appointment. Friends who note “Name – Company/School – Position/Major” can be given priority for approval.
Course Reminder
The 4 lectures of the Embedded AI series have been live-streamed on the Zhidx public course mini-program and support course reminder reservations. Friends who are worried about missing the live broadcast can click the “Read Original” at the bottom of the article to enter the live broadcast room, and click the “Remind Me When It Starts” button to subscribe. You will receive a notification 5 minutes before the live broadcast begins.
Give it a “look” and chat with everyone
👇👇👇