SEOUL, August 11 (AJP) - A self-driving car has to work out where every other vehicle, pedestrian and obstacle around it sits in three-dimensional space, and it has to do so many times a second on a computer bolted inside the car rather than in a data center. A South Korean university team says it has found a way to shrink the software that does the recognizing without losing accuracy.
Kookmin University said Tuesday that two master's students in its artificial intelligence and software department, working with Hyundai Motor Company, had developed a compression method called SharedKD for the three-dimensional object detection models used in autonomous driving. Cho Hyun-jun and Ahn Sang-ho carried out the work in the research group of Kim Jang-ho, and the paper was presented in July at the Design Automation Conference in Long Beach, California.
The problem the method addresses is a squeeze between two demands that pull in opposite directions. Detection models grow more accurate as they grow larger, but a car cannot carry a rack of servers, and a model that takes an extra fraction of a second to produce an answer is a model that has seen a hazard too late. Anything that runs on a vehicle has to fit within a fixed budget of processing power, electricity and heat.
The standard fix is a training technique called knowledge distillation. Engineers take a large, accurate model, call it the teacher, and use its outputs to train a much smaller model, the student, so the small one learns to imitate the big one's judgments rather than learning from raw data alone. The result is a model that runs on modest hardware while behaving more like its oversized parent than its size would suggest. Distillation normally requires building and training two separate models.
SharedKD does away with the second one. The full network serves as the teacher, and a subnetwork carved out of that same network by pruning, which means switching off connections judged to be doing little work, serves as the student. During training, the system uses gradients, the signals that tell a neural network how much each part of it is contributing to the final answer, to decide which parts to keep. The choice is made continuously as training proceeds rather than fixed in advance.
The practical difference is one of sequence. Compression is usually something done to a model after it has finished training, a trimming of a completed object. SharedKD looks for the deployable small model at the same time the large one is being trained, and the two are set up so that each improves the other's learning. Kookmin said the approach cuts both the amount of arithmetic a detection model performs and the time it takes to produce a result, and expects it to be applied in vehicle systems and in edge AI, the general term for artificial intelligence that runs on a device rather than sending data away to be processed.
How much it cuts has not been made public. The university released no performance figures for SharedKD, named no baseline model or test dataset for comparison, and did not say whether the method has been run on automotive hardware or only in the laboratory. Neither Kookmin nor Hyundai has described any plan to put it in a production vehicle.
Model compression is Kim's own specialty. He joined Kookmin in September 2022 from Naver Webtoon's AI Research Lab, where he worked on meta-learning, and his doctoral research dealt with the tendency of deep learning models to become unwieldy as they become more capable.
Kookmin said the project reflects its KMU Vision 2035: EDGE strategy, which designates eight areas for concentration, and that the work sits where two of them meet, the AI+X track and the mobility track.
The Design Automation Conference (DAC) is the main annual gathering for the design of electronic circuits and systems, sponsored by the Association for Computing Machinery and the Institute of Electrical and Electronics Engineers. This year's edition, the 63rd, ran from July 26 to 29 at the Long Beach Convention Center, and organizers said more than half of the program dealt with artificial intelligence.
Copyright ⓒ Aju Press All rights reserved.