AlphaGo computer victory sparks AI ‘deep learning’ drive
The strategic game of Go was widely viewed as one of the major unsolved 'grand challenges' for artificial intelligence after decades of AI research.
Go grandmaster Lee’s defeat in all but one game with the computer programme developed by Google has been a wake-up call, particularly in the field of ‘deep learning’, where the programme learns from available big data and from its own mistakes.
In the wake of the computer’s win, the South Korean government announced it would invest US$863 million in AI research over the next five years, including bringing forward the opening of a previously announced public-private research centre involving a number of Korean chaebol or industrial conglomerates such as Samsung and Hyundai.
South Korean President Park Geun-hye said in a speech on 17 March that AI was “the fourth industrial revolution”.
“Korean society is ironically lucky that thanks to the AlphaGo shock we have learned the importance of AI before it is too late,” she said.
South Korea’s AI research institute already planned for 2017 will now open even earlier. “Once the private companies set up the research institute, the government will provide financial support for core R&D projects,” the ministry of science said in a statement.
Tentatively named the Intelligence Information Technology Research Institute, it is expected to be in Pankyo, just south of Seoul, according to Kim Yong-soo, a senior ministry official.
Catch up with the West
Japan in particular fears it has fallen behind the US and other countries in the field of ‘deep learning’. The government established a new AI research centre in May last year at the National Institute of Advanced Industrial Science and Technology and is planning a second this year in order to break out of the more narrow focus in Japanese research of AI applications to robotics.
Japan’s 5th Science and Technology Basic Plan 2016-2020 by the cabinet of Prime Minister Shinzo Abe in February includes AI as a priority area for research.
The drive to catch up swiftly with the West in this branch of AI has meant luring ethnic East Asian talent already working on such projects in and setting up research labs in California’s Silicon Valley to tap into the best talent.
Top brains are being routinely lured by Asian companies’ research departments from Google, Facebook and IBM which have invested heavily in AI.
Japanese car giant Toyota announced late last year it would hire some 200 researchers in a five-year, US$1 billion research institute in collaboration with Stanford University and Massachusetts Institute of Technology, expected to be one of the largest research laboratories in Silicon Valley mainly focusing on applications of ‘deep learning’ to steer safer driverless cars.
China’s new ‘brain’
In China, Robin Li Yanhong, founder and chief executive of Chinese online search engine Baidu, said at a press conference in Beijing in March 2015 that he was looking to secure Chinese state funding – including from the military if necessary – for a massive AI “China Brain” project which he compared to the US Apollo space mission that landed a man on the moon in 1969, the human genome project and other big science projects.
“This [AI] is one of the peaks of world technology, and China is not lagging behind in this area, so I think we have an opportunity to do something big,” he said.
AI has been listed in China's 13th Five-Year Plan 2016-2020, as a “highly valued” sector that should be developed as a key strategic project. But so far specific sums for AI research have not been officially announced.
Meanwhile Baidu set up its ‘deep learning’ research centre in Silicon Valley in 2014 investing a reported US$300 million in the lab.
Late last year it announced its programme had surpassed humans in recognising Mandarin and English for the first time. The ‘Deep Speech 2’ system learned to recognise words by listening to many thousands of hours of transcribed audio. The breakthrough will allow more accurate voice queries for Baidu searches, the company said.
Baidu’s chief scientist Andrew Ng, a former professor at Stanford University, also known for co-founding the Coursera massive open online course, or MOOC, platform, said in December that for short out of context phrases, the programme appeared to be surpassing human levels of recognition.
“It’s the same as the way a baby would learn: we show [the computer] audio, we show it text, and we let it figure out its own mapping...” he said in a recent interview.
Last year, 2015, has been described by AI experts as a breakthrough year for ‘deep learning’ programmes in Asia, not just AlphaGo.
The University of Science and Technology of China announced last year it had taught a computer network how to take an IQ test and said it scored better than a typical postgraduate student.
China is also pitching computers against humans in its competitive gaokao university entrance exam.
“Our goal is to make a robot smart enough to enter a first-class Chinese university in 3 to 5 years,” Liu Qingfeng president of tech firm iFLYTEK Co, was quoted by China’s official Xinhua news agency as saying in December.
“It is easy for a robot to sit the gaokao because machines are strong in memory,” Liu said, “but it is hard for them to surpass 80% of human candidates to qualify for first-class universities.”
China’s Guangdong and Guangxi provinces already use the company’s system to replace teacher marking in the gaokao English oral test. The company’s aim is for an AI robot to take part in the 2020 gaokao and be in the top 20% of scorers – some 9 million students take the gaokao across China every year.
“AI’s participation in the gaokao is the best proof of its ability to learn and reason,” Liu said.
Japan’s National Institute of Informatics announced in November its Todai AI programme, jointly developed with university and corporate researchers, had for the first time achieved an above-average score in the country’s university entrance exam in five subjects including mathematics, physics and English.
The AI programme scored 511 points out of 950 – above the national average of 416, and did exceptionally well on mathematics and history-related problems, the institute said. With that score, the AI has at least an 80% chance of being accepted to 441 private universities and 33 national universities, according to the institute.
The software programme had been cramming for the exam since 2011 but had below-average scores on similar exams in 2013 and 2014.
The institute has been working on an AI programme able to score enough on Japan’s standardised college entrance exam to be accepted into the University of Tokyo, the nation’s top university. That goal has now received added impetus from the success of AlphaGo.