Author: Tronserve admin
Monday 2nd August 2021 10:59 PM
Megvii Makes Deep Learning AI Framework Open-Source
Chinese artificial intelligence (AI) start-up Megvii Technology announced on Wednesday that it would make its deep learning framework open-source, as China steps up the development of home-grown AI and makes the technologies more accessible.
Initially developed in 2014, MegEngine is part of Megvii’s proprietary AI platform, Brain++, which can train computer vision at large scale and enable developers across the world to build AI solutions for industrial and commercial use, according to the Beijing-based company.
“If we liken the process of algorithms creation to cooking a meal, then data is the raw material and computing power the stove fire,” explained Tang Wenbin, Megvii’s co-founder and chief technology officer at a live-streamed launch event on Wednesday. “Those algorithms would still need a framework to carry them and MegEngine is the pot.”
China’s reliance on US-originated frameworks is seen as a significant gap in its AI ecosystem: despite an ambitious plan to become a global leader in AI by 2030, nearly all small- to mid-sized Chinese AI companies rely on open-source platforms from the US which offer a host of tools and libraries designed for machine learning and deep learning techniques that teach computers to learn by example.
“The latest tide of AI cannot live without deep learning technologies, whose development has everything to do with open-sourced infrastructures,” said Gao Wen, a Peking University professor who is also director-general of the country’s New Generation AI Technology Innovation and Strategic Alliance.
Established open-source platforms like TensorFlow and Pytorch – both from the US – democratise deep learning, allowing almost anyone to feed data into these models and start training their own AI systems without having to create their own from scratch.
Amid the ongoing tech war with the US, concerns that China’s relative weakness in foundational AI infrastructure such as algorithms and frameworks may derail its ambitions in the field have been brewing.
Megvii was put on a US trade blacklist last year, along with seven other Chinese AI companies, over alleged human rights violations.
The company, which was preparing for a Hong Kong initial public offering, said last October that the decision by the US government could disrupt its supply of servers as well as its IPO plans, but that it was “ready for the fight”.
MegEngine’s Chinese name, Tianyuan, means “the origin for everything” and also refers to the centre point of a Go board. The name is not only a salute to AlphaGo, the computer program developed by Google’s DeepMind, but represents Megvii’s wish to enable a better future for the country’s AI industry, Tang said.
“China has been taking part in open-source since the Android system was introduced in 2011, and Chinese developers today make up an increasingly large proportion of the global open-source developer community,” Tang said.
“We think that open-sourcing MegEngine is consistent with this trend. Megvii wants to lead the way in China and create goodwill among developers here and abroad.”
Still, China is playing catch-up with the US, as building an in-house framework from scratch requires a large investment of time and dedicated resources, as well as top-tier talent, to ensure the framework runs smoothly and covers a variety of use-cases.
Tensor Flow and PyTorch hold a combined 95 per cent of market share for deep learning frameworks, Gao pointed out.
Chinese search giant Baidu introduced its PaddlePaddle open-source AI platform in 2016, only a year after TensorFlow was launched, but it failed to gain traction among global AI programmers.
Megvii said in its news release that is “one of the very few companies in the world that has developed its own deep learning framework”, and that by making parts of it available for free in the alpha launch it hoped external developers both in and outside China would be able to improve MegEngine for the benefit of all users.
The framework, which contains 300,000 lines of code in its alpha version, is available for downloads at the company’s website and Github, it said. Megvii has a beta version tentatively set to launch in June and expects to launch the official version in September, it added.
Last week, a research team in the Department of Computer Science and Technology at Tsinghua University also open-sourced its deep learning framework Jittor. Based on Just In Time compiling and meta-operators, researchers said the framework is easy to use because it is composed of a basic layer and users can easily generate customised high-performance code for the models.