Organizing Memos from Recent Studies

Organizing Memos from Recent Studies

-Physical Embedded AI market is much larger than existing Cyber/On-device AI market

-The traditional autonomous driving method can easily achieve 80% of the functionality with 20% effort. However, continuous improvement is not possible. (Dr. Xinzhou Wu, former head of Xpeng Autonomous Driving Division, 現 Nvidia Autonomous Driving Head)

-While autonomous driving by deep learning has a low and high upper bound, traditional autonomous driving has a high and low lower bound. (Xudong Cao, Momenta CEO)

• In the case of dev-and-conquer, practical solutions can be reached quickly with 20% effort, but the more the components are split, the more likely they are to be trapped in the local optima. If you reduce the performance in stages to escape from the local optima, you will be trapped in the 80% performance upper limit. Therefore, an end-to-end method that considers all elements at once (to reach the global optima) is more suitable for smooth autonomous driving. Divide-and-Conquer uses a modular approach, and End-to-End uses an integrated approach. For this, the end-to-end method has no choice but to rely on more data. The difficulty of interpretation is generally more easily interpreted by the divide-and-conquer method. Divide-and-Conquer can optimize each module individually, but End-to-End must optimize the entire system at once.

-Among self-driving companies, only 3-4 Tier 1 companies will survive, possibly with a 7:2:1 control. The No. 1 company will control 70% of the global market

-End to end AI Architecture will ultimately win. A technology that clearly confirmed that the data-driven approach was correct in the end, became more pronounced when transformer algorithms emerged in 2017 and began to be integrated into the field of Vision

-Competition will intensify as companies like Nvidia and Horizon jump directly into self-driving

-Based on China in the first half of 2023, Nvidia and Horizon Robotics account for 80% of the high-end Navigate On Autopilot (NOA) platform. Nvidia accounts for 52.57% and Horizon 30.71%, with Texas Instruments (TI), Mobileye, Huawei, and others occupying the rest.

-Tesla’s easy integration of the End to End AI Architecture was thanks to Andreekapaci’s outstanding artificial intelligence design. Tesla’s Shadow Mode is a key source of Tesla’s training mode, and is designed to trigger the data return mechanism when the path selected by the system deviates significantly from the driver’s choice. The system is designed to automatically store camera capture in memory at the time, and update received data and vehicle driving data to Tesla servers after connecting to Wi-Fi. (Data engine)

-Ashok then took over Andrekapasi and introduced the Occupancy network. The video captured by the angle camera is restored to a 3D scene. AI calculates how much space an object occupies. It is a method of inferring the shape by looking at the points that an object occupies in space. As a result, LiDAR and radar are not needed.

-In the case of Tesla End to End AI Architecture, all over 300,000 lines of rule code coded by engineers have been removed. The initial 1 million video clips were not enough, but the improvement increased as it exceeded 2 million clips, and the name was all admired as it exceeded 3 million. AI’s driving skills become incredibly natural as it crosses 10 million clips

-Embodied AI is about how to solve changes, complexities, and exceptions to the physical world. In addition to data and experiences, it has evolved by using ① Meta Learning, ② Transfer Learning, ③ Causal Learning, and ④ Active Learning, 방식 How AI Systems Identify and Learn Where Information Is Most Informed on Their Own. ⑤ Human-AI Cooperation Model exists. It can be advanced by combining human knowledge and AI capabilities, which necessarily requires high-quality human knowledge

-For hardware in the autonomous driving industry, Moore’s Law, which halves prices every two years due to the rapid evolution of artificial intelligence semiconductors, will be applied, and software will evolve 10-100 times every two years with the help of artificial intelligence. Therefore, it is highly likely that a small number of companies that have achieved Vertical Integration that internalizes both hardware and software will win the competition and dominate the global market. Businesses that have as high technological barriers as autonomous driving and have large economic effects of scale will also be implemented (Xudong Cao, CEO of Momenta)

-China’s smartphone business has developed significantly since Apple’s production in China, with Oppo, Vivo and Xiaomi seeing significant improvements in their competitiveness. Similarly, electric vehicles have developed remarkably since Tesla’s Shanghai factory operation, and Nio, Xpeng, Li Auto, BYD, and Xiaomi have been able to appear. Once Tesla’s FSD is allowed in China, competition in self-driving will intensify, but Huawei, Xpeng, Momenta, Weiride, and PonyAI are expected to grow significantly after Tesla. (Xudong Cao, CEO of Momenta)

-Full self-driving is a difficult technology that requires solving the long-tail risk of 1/10,000 and 1/100,000. Therefore, it is almost impossible to find and improve the edge case and corner case of full self-driving with 1,000 test cars