HANGZHOU, CHINA--Alibaba Group has formally established a semiconductor business to produce its own artificial intelligence (AI) chip as well as unveiled plans to develop quantum processors.
Driven by the Chinese vendor's research and development (R&D) arm Damo Academy, these efforts would see the launch of Alibaba's first in-house developed AI chip in the second half of next year. Called AliNPU, the new AI chip had the potential to support technologies used in autonomous driving, smart cities, and smart logistics, said Alibaba at its annual flagship computing conference held here Wednesday.
It also set up a new semiconductor subsidiary, called Pingtouge, which it said would focus on customised AI chips and embedded processors. These efforts would support Alibaba's plans to expand its cloud and Internet of Things (IoT) businesses as well as drive the development of industry-specific applications, it said.
Alibaba in April acquired integrated circuit design vendor, Hangzhou C-Sky Microsystems, describing the move as "an important step" in boosting its chip-making capabilities. The move also would marry both companies' R&D strengths and was in line with China's urging for the country to become self-reliance in the development of key technologies.
It then had revealed initial plans for AliNPU, which it said would be designed to process AI tasks such as image and video analysis.
Alibaba CTO Jeff Zhang said at the conference: "Moving ahead, we are confident our advantages in algorithm, data intelligence, computing power, and domain knowledge on the back of Alibaba's diverse ecosystem will put us at a unique position to lead real technology breakthroughs in disruptive areas, such as quantum and chip technology."
In laying out its five-year roadmap for Damo Academy, Alibaba said the R&D arm also would be developing "high-precision, multiple-qubit superconducting quantum processors" and would continue to push development in the sector. These would include quantum-classical systems to offer utility-based quantum compute power that could be delivered over the cloud.
The academy also would look to expand its partner ecosystem to identify potential quantum-powered applications for various industry verticals, including logistics, e-commerce, materials, and pharmaceuticals.
Zhang said the development of both software as well as hardware was necessary to provide the computing necessary to more quickly analyse data and at a low cost.
"Computing begins with chip. We established a team last week [and] by mid-2019, we will have the first neuro network chip," he said, adding that the AliNPU had shown--in current tests--to increase image processing performance by four-times.
Google also offers its own AI chip, first announced in 2016 and currently in its third generation. The Tensor Processing Unit (TPU) 3.0 is touted to be eight-times more powerful than its predecessor.
Launched last year, Damo Academy currently has more than 300 researchers across eight cities worldwide, focusing on five technology areas including fintech, robotics, and quantum computing. Its partners include Nanyang Technological University of Singapore and Stanford University.
Future AI development should explore combination of senses
Future development of AI technologies could explore the use of multiple "senses" to provide more accurate predictions or analysis, said Rong Jin, Damo Academy's head of machine intelligence technology.
Speaking to media at the conference, Rong explained that most AI software development typically focused on just one technology area. Speech recognition technology, for example, would focus only on speech, while visual recognition software would look only at visual technology.
Little thought was given to combine the analysis of different senses, such as visual and speech, he said, noting that humans were able to multi-task because they would use multiple senses at the same time. This enabled them to gain a bigger picture and better understanding of the situation.
He pointed to Alibaba's collaboration with Shanghai Shentong Metro Group to develop a system that could enable passengers to purchase subway tickets using voice interaction. The station's open environment, however, was likely too noisy and technically challenging for the system to accurately recognise voice.
Visual recognition technology then was used to identify the speaker's lip movements or lip-read. These visual signals were used together with audio signals to improve the interpretation of the customer's ticketing order.
This collective use of multiple technologies could be used in AI to establish a better understanding of real-world challenges, Rong said.
Another potential development of AI could explore humans' ability to learn based on a small database, compared to the copious amount of data on which machines needed to train, he noted. For instance, machine translation technology today had made significant progress, but needed to train on millions of books to do so. In comparison, humans were unlikely to have the capacity to do likewise, but most would have grasped a good understanding of a language by reading a couple of hundreds of books, he noted.
"That's a huge contrast...so how do we bridge this gap? And how can we enable machines to learn new skills with [a smaller sample database]?" he posed, adding that future AI development could look at how the technology could be more like humans in learning.
Based in Singapore, Eileen Yu reported for ZDNet from The Computing Conference 2018 in Hangzhou, China, on the invitation of Alibaba Group.