By now, you've likely heard about AlphaGo's 4-1 triumph over a human expert at Go, an ancient game so complicated that many experts in the field had projected such a win was still a decade away. We're inundated with new scientific findings and 'progress' on a daily basis, and at times it can be difficult to evoke appropriate wonder; to put some perspective on the extent of this feat, Google's Demis Hassabis likened the win to landing on the moon - a milestone for machine learning and the field of AI.
Google DeepMind's software program symbolizes a real leap in the evolution of AI's capabilities, and implications of the software's potential are already beginning to ripple across industries, with new doors seeming to swing open in finance, business, manufacturing, and others.
The world saw the same reaction with Watson when it beat Gary Kasparov at Jeopardy in 1997, though it's only been within the last two years that IBM Watson opened up to enterprise. The technology platform now boasts close to 300 (and growing) commercial partners, along with billions of API requests monthly from entrepreneurs to tech enthusiasts. It's only a matter of time, and I would venture on a much shorter time scale, that Google DeepMind's technology follows suit.
An email exchange from the Future of Humanity Institute unearthed thoughts from well-known Futurist Nick Bostrom, who remarked that "AlphaGo is more interesting than either Deep Blue or Watson, because the algorithms it uses are potentially more general-purpose. It wouldn't surprise me if, with very minor modification, the same approach will perform well on a wide variety of perfect information games."
Perfect information is a key concept in making sense of AlphaGo's capabilities and its near-term impact in industry within the next decade. Perfect information defines a game in which all moves are made based on having complete knowledge of all previous moves made by another player(s). This would rule out, of course, any game in which moves by various players are made simultaneously or out of view of another player. Many real-life business scenarios, including making important strategic decisions, often fall into the imperfect information category, and these arenas - to the relief of many - are still out of AI's hands (for the time being).
Where AlphaGo has real potential to disrupt industry soon is in making lots of smaller, sequential decisions on a continuous basis, as well as solving problems quickly based on specific parameters. Deep learning also has a talent for spotting anomalies amidst large sets of data, and its ability to understand natural language in context - dubbed "sentimental analysis" - is also continually being refined.
In a young industry, there's lots of room for creativity and inevitably there will be surprises, but companies are already tapping deep learning for tasks related to marketing campaigns, lead conversions, email flows, interactions with corporate teams and clients, and more. In a personal email exchange, Silicon Valley Entrepreneur Steve Omohundro remarked that deep reinforcement learning networks are already being used for autonomous drone control, robot bin picking, and autonomous vehicle control. From use in image recognition to doling out customized advertisements, deep learning is a quickly advancing and lucrative field, with Gartner predicting that 20 percent of business content will be "machine authored" by 2018.
In addition to heavy investments from mammoths like Google and Facebook, smaller companies are also hopping on board. East-coast based ADP, for example, is using deep learning systems to help scout out necessary information, run big data analysis, and present prepared report to its CEO, a process that is improved upon each time the machine performs its operations. Saffron, a division of Intel, is using deep learning to match broad patterns of customer behavior to specific individuals, claiming that the technology predicts correct next moves - such as how the person will contact a company - 88 percent of the time. Companies such as Rare Mile Technologies are also offering up customized machine learning algorithms to a range of industries for various uses, including insurance fraud detection.
An important difference between traditional machine learning and AlphaGo's breed of deep learning is the increasing complexity of problems that the technology is being employed to solve, which requires deep analysis combined with the ability to build upon layers of past experience, ultimately presenting better decisions and solutions over time.
As described in a paper authored by Google researchers and published in the November 2015 edition of Nature, AlphaGo is different from past systems in its use of two distinct but cooperative deep neural networks. The first network figures out which moves are a best fit for the situation at hand (described as reducing the "width of the search space"), and the other network learns to infer desired outcomes based on a particular move (reducing the "depth of the search space").
Researcher Aubrey de Grey commented on AlphaGo's outcome, stating that "the breakthrough that DeepMind has made - not simply by combining deep learning with reinforcement learning, but by enhancing that combination in the form of their so-called "deep Q networks" - can, I believe, lay claim to being the single biggest advance in the entire history of artificial intelligence research. The fact that essentially the same algorithm can master a variety of video games and also the most daunting of all board games tells us that its generality is very great."
Of course, creating AI systems that play games is easier than creating them for imperfect scenarios, where the data and moves get "messy" - basically, any real-world context, from business to healthcare. Not to mention that it's much more difficult to train networks to continuously learn from situations in which people's lives and welfares are at stake. In the game of Go, despite the inconceivable number of possible moves, the rules and observable information are all pre-determined.
Yet one of the most interesting outcomes of the tournament was AlphaGo's ability to consider moves and options that are not known or are usually ignored by human beings. The AI made "surprising" moves that a typical human professional would not consider. In a very real sense of the definition, AI is creative, which can potentially be used to help humans solve increasingly complex problems on an organizational or a global-level scale.
Researchers will doubtless start applying deep mind technology toward imperfect information games, such as Poker, in which machines will try and read human actions as clues and infer the best move without having all of the other player's information beforehand. Early incomplete information applications might include certain types of business negotiations or in cybersecurity.
Success in enterprise requires a plethora of mundane tasks. If a system can tackle much of the important research and groundwork involved before an individual or team goes to make decisions, all the while learning the difference between recognizing critical observations and irrelevant anomalies, humans will be given the opportunity to spend more time honing in and refining our abilities in those areas in which we still have the upper hand - perception, long-term visioning, and the cultivating of better human relationships.
AlphaGo's win seems to reveal the dawn of a promising new era, one in which humans have a valuable opportunity to team up with AI to solve problems, reduce errors, and operate more efficient systems across industries.