The race to build artificial intelligence (AI) computing chips ramps up. Last month Google GOOGL -2.77% revealed the development of its new processor Tensor Processing Unit to accelerate this field and the chip will be used at its data centers. Meanwhile Chinese company Horizon Robotics shared similar ambitions at EmTech Hong Kong tech conference last week. Dr. Kai Yu, founder and CEO of the company said the firm is developing “cutting edge computer infrastructure to facilitate AI computing, perception, cognition and real-time decision-making on smart devices.”

Yu was formerly head of Baidu BIDU -1.81%’s Institute of Deep Learning. “We are developing neural networks and software solutions and hardware together [in the hope to] bring more AI to our lives,” he said.

Neural networks, also referred to as deep learning, are software techniques that emulate human brain function; algorisms in a multi-layered web of virtual neurons mine through large volumes of data to look for patterns, learn from those patterns then make decisions intelligently. Many tech giants employ this technology, which has advanced image and speech recognition abilities in machines and online services. That includes Facebook FB -1.26%’s facial recognition features, which can identify people in photos uploaded onto the site.

Yu says there are limitations of conventional computers today, which are powered by Intel INTC +0.16%’s chips, “This general computing infrastructure uses a serial computing structure, which means every task is done [one step at a time],” he explains. Whereas a new generation processor specifically for A.I. would be much more efficient and designed for cognition and perception to enable decision-making in machines.

The underlying systems governing these new processors will facilitate computing through neural network type of computation, he says. “We are also developing the software platform on top of this computing architecture,” he says. The key focus areas to apply such innovations will be on home appliances and self-driving cars.

At the conference, Yu showcased Hugo, its newest software, a smart car system to enhance safety on the road for self-driving vehicles. The system can detect vehicles, pedestrians among other objects and obstacles on the road simultaneously, thanks to its underlying deep learning algorisms or “neural network” platform.

Asked when autonomous vehicles will be ubiquitous on roads, Yu predicts this future will be a while away. “Even though I’m working on this, I believe we have 15 years to go,” he says. A critical hurdle is developing software for such vehicles. These cars are equipped with sensors and other technologies to detect and respond to objects around them in real-time and anticipate what objects will do next. However, self-driving software are tested in advance and items on the road are compiled, based on real world scenes, so vehicles can classify and react to these cases on the road. But streets are not always predictable; what if new object or obstacle pops up on the road? “It’s a real challenge and you don’t know how to test it because you don’t know what kind of case [will appear that you have never seen before].”

By: DocMemory
Copyright © 2023 CST, Inc. All Rights Reserved