+ A

A world where AI has an imagination

Google DeepMind co-founder discusses AlphaGo’s successors
Apr 02,2018
이미지뷰
Demis Hassabis, CEO of Google’s DeepMind, poses with late physicist Stephen Hawking. Hassabis, Hawking and Tesla CEO Elon Musk all endorsed a set of guidelines last year for ethical AI development that will benefit humanity. [DEMIS HASSABIS]
Two years have passed since the legendary match between Google DeepMind’s AlphaGo artificial intelligence program and Korean Go master Lee Sedol sent shockwaves across Korea. AI has been a buzzword in Korean society ever since. As AlphaGo retired from competitive gaming in March 2017, the company has instead been concentrating on tackling “a very wide range of problems” that humans find difficult to resolve.

“The majority of the AlphaGo team now devotes their time to new projects with the intention of using these general-purpose algorithms to help solve some of the world’s most complex challenges in science and medicine,” said Demis Hassabis, co-founder of DeepMind, in an email interview with the JoongAng Ilbo to mark the two-year anniversary of the high-profile match.

Projects the Google-owned British start-up is now devoted to range from developing an AI program to diagnose eye diseases to the more abstract task of creating an AI with an imagination, according to the chess prodigy-turned-entrepreneur.

“We believe a similar ability to plan for the future will be hugely beneficial for AI and we have made important steps to incorporate this ability into our artificial models,” he said.

Despite the progress of other AI-based Go software developed in many other countries in the past two years, DeepMind has no plans to take AlphaGo out of retirement to face off against the newer challenges, Hassabis noted.

“We have been really happy to see other teams taking inspiration from our work to build their own systems,” he said.

Hassabis went on to imagine a future where AI is routinely used as a tool to help scientists make discoveries and complete experiments far faster than would ever have been possible without the computer programs.

The following are edited excerpts from the interview.



Q. DeepMind has been developing an AI program to diagnose eye diseases. Could you explain more?

A.
Clinicians at the world renowned Moorfields Eye Hospital in London contacted us because they were keen to explore whether AI could help speed up the diagnosis and treatment of major causes of sight loss. They told us the volume of scans they need to analyze per week puts strain on the system and can lead to delays in treatment, which is a serious problem for people with deteriorating sight loss. Now we are working with the team there on a research study to analyze anonymized retinal scans. It’s still in its early days and we can’t pre-judge the results, but we believe this could be the start of more effective, faster treatment of sight loss.



In what areas can DeepMind’s AI creations be applied to?

We’re really excited about applying AI to healthcare and we’re currently conducting a few research studies. There’s our study with Moorfields, which I mentioned earlier, and we’ve also recently announced a new partnership with the Cancer Research UK Centre at Imperial College in London to help detect breast cancer with machine learning. We’re hopeful this will enable expert clinicians to better triage the thousands of scans they process per week and signal the start of a more effective, faster treatment of this type of cancer.



DeepMind announced earlier that it is developing AI capable of “imagination.” Can an imaginative AI create and plan for the future?

Imagination is a hugely important function for humans, allowing us to think through the consequences of our actions before we take them. This saves both time and effort. We believe a similar ability to plan for the future will be hugely beneficial for AI and we have made important steps to incorporate this ability into our artificial models.



What is DeepMind’s current status on developing general-purpose algorithms?

Our intention is to build general-purpose algorithms that can be applied to a very wide range of problems. Our most recent example was AlphaZero - a more general version of AlphaGo, which is able to play chess, shogi and Go better than any other program.

The movie “AlphaGo” was released on Netflix in January. It depicts the match where AlphaGo won four out of five matches against Korean Go master Lee Sedol. What do you remember the most about the match?

It was an amazing, once in a lifetime experience that completely surpassed any of our expectations. What really stood out for me was the warm welcome we had from the Go community in Seoul - from the incredible support of the Korea Baduk Association through to the generous sportsmanship of Lee Sedol himself. We were amazed by the level of interest in the tournament from people all around the world and we’ve been thrilled to see the game of Go soar in popularity in western countries as a result.



AlphaGo retired from the Go game but new Go AI programs such as Chinese tech giant Tencent’s Jueyi, Japan’s DeepZenGo, Taiwan’s CGI and Korea’s Dolbaram have all been improving. Does AlphaGo have any plans to come back and play a match with another Go AI or professional human player?

We have no plans to bring AlphaGo out of retirement. Our matches in Seoul and in Wuzhen were the highest pinnacle for AlphaGo as a competitive program, and it was a real thrill and honor to have that experience.

We took the decision to openly publish our research for AlphaGo and AlphaGo Zero, with all of the technical details, as a gift to the Go world, and we have been really happy to see other teams taking inspiration from our work to build their own systems.



DeepMind is working with Blizzard Entertainment to develop an AI program that can beat human players in the multiplayer videogame StarCraft II. Are there updates on the progress?

StarCraft II is a great environment for training and testing our algorithms because it emphasizes capabilities such as planning, imagination and strategy. Our focus is on fundamental research in the StarCraft II environment right now, but it’s an exciting idea!



What other AI programs are you interested in developing?

As a chess player, I have been really amazed by the completely new style of chess AlphaZero plays, favoring mobility over material. We hope to be able to share more about this in the near future.



What will DeepMind be doing 10 years from now?

We’re nearly at the point where we’re able to start applying our learning algorithms to science itself. I look forward to making our first breakthrough in a really challenging area of science and showing that AI can be used as a tool to help experts in their field make discoveries that would otherwise have been impossible. I hope we’ll start seeing this happen in the next couple of years, and then by 10 years it will hopefully be routine.



What is DeepMind Ethics & Society?

When you’re developing technologies such as AI that could have a transformative impact on society, it’s important to engage with broader society about the development and deployment of your work. We set up a dedicated research unit called DeepMind Ethics and Society, which works with a whole bunch of academics including people like Professor Nick Bostrom from the Future of Humanity Institute at Oxford to explore the potential impact of AI, as well as its ethical deployment and use. We started this unit with external experts like this because we don’t have all the answers - no one does - but we have got to start thinking about these issues seriously and having companies, academics and nonprofits come together to think about them. These are huge challenges on the whole spectrum of things from technical safety to ethics and policy, and I am personally invested in making sure AI is deployed safely and that productivity gains benefit as many people as possible.



How do you work with Google?

DeepMind operates independently, as it has done since we were acquired in 2014, but we work with Google on many projects. For example, in 2016 we applied our general purpose learning techniques to help improve the efficiency of Google’s already highly-optimized data centers. Our system was able to make recommendations which led to cooling energy reductions of up to 40 percent and an overall 15 percent improvement in the data centers’ energy efficiency.


BY SOHN HAE-YONG [seo.jieun@joongang.co.kr]