Ilya Sutskever, director of OpenAI, an independent research group, will describe what might be the next big breakthrough in artificial intelligence today at EmTech Digital, a conference organized by MIT Technology Reviewin San Francisco.
Sutskever will describe research showing an approach in machine learning that can perform even better than methods that have produced huge breakthroughs recently. His technique may also prove far more scalable.
In a blog post describing the work, Sutskever and colleagues describe using “evolutionary strategies” to have machines figure out for themselves how to solve a complex task. The researchers say the approach is distantly related to a decades-old approach that involves optimizing algorithms using a process of simulated evolution. It essentially lets a machine work out, using experimentation and optimization, the best algorithm for solving a complex problem, and it could have applications in robotics, automated driving, and other areas.
The OpenAI researchers compare their evolutionary strategies approach to reinforcement learning, a technique that has produced some impressive results in the past year or so, including enabling a computer to defeat one of the world’s best Go players (see “10 Breakthrough Technologies: Reinforcement Learning”). Reinforcement learning, which is loosely based on the way animals seem to learn through experience, enables machines to figure out how to do things that are difficult or impossible for a person to describe in code.
Unlike reinforcement learning, evolutionary strategies allow machines to learn while using much less computation. Reinforcement learning typically requires a technique known as backpropagation, which optimizes a neural network as errors are minimized. Evolutionary strategies involve a much simpler optimization technique.
“This is very interesting and could indeed be the start of something larger,” says Pedro Domingos, a professor at the University of Washington and the author of The Master Algorithm, a book about different machine-learning methods.
Domingos questions whether the technique will surpass reinforcement learning, but he adds: “There is a delightful history in machine learning of very simple methods coming along and beating much more complex ones. It's about time we saw a broadening of approaches.”
Big technical breakthroughs will be needed to sustain the momentum seen over the past few years in artificial intelligence.
Having upended the tech world in recent years, artificial intelligence now seems poised to find important uses in industries including health care, education, transportation, and manufacturing. For instance, deep learning, a technique that involves using very large neural networks to find patterns in data, and which has proved to be a very powerful voice- and image-recognition technique, is rapidly finding applications in medical research and health care.
At the same time, however, it has become clear that these technologies alone will not provide the general artificial intelligence that has long been the dream of the field. The tension between new opportunities and the continued need for innovations will be prominent themes at EmTech Digital.
Even with the emergence of new machine-learning techniques, such as reinforcement learning and OpenAI’s evolutionary strategies approach, the ultimate goal of the field—some form of artificial general intelligence—remains a distant vision.
Still, the spread of powerful machine learning into new industries and areas of daily life will heighten attention on the unintended consequences that may result.
Speakers at EmTech Digital will discuss the issue of the bias that can become embedded in machine-learning algorithms that are increasingly used to guide important decisions such as the appropriate length of a sentence for a person convicted of a crime, or who is granted a bank loan.
As machine learning is deployed in more areas of life, this issue will become more important, and will raise serious ethical concerns. The difficulty of interrogating the latest machine-learning algorithms to find out how they made a decision could compound this issue, and will be another theme at Emtech Digital.