Google’s AlphaGo, IBM’s Watson and Predictions from the MIT AI Conference

5979

This week a computer finally managed to beat the top human exponent of the Japanese game of Go. It has been over 20 years since world champion Gary Kasparov lost a single game to a computer in 1996. A year later he lost an entire series. This week world champion Go player Lee Sedol lost his first game, the first of a series of 5, to the AlphaGo program running on Google DeepMind.

Go is based on placing stones on a 19 by 19 grid, and so a board has 361 factorial moves possible, which amounts to infinitely more potential moves than chess. It has only been the resurgence of Artificial Intelligence, as a viable technology, and the open sourcing of key ingredients, that has allowed this leap from the complexities of chess to those of Go.

All this means is that after almost 50 years of false starts, Artificial Intelligence (AI) is breathing new life into nearly every industry – transportation, retail, fashion, medicine, finance, recreation, communication, and more. We went to MIT’s artificial intelligence conference in Boston last month, and realized that this is an industry on the verge of rebirth.

AI, interchangeably used with machine intelligence, is a term for methods, algorithms, and technologies that make software smart like a human. Retailers anticipate a big year in 2016 as they inhale customer shopping behaviors and preferences using cognitive computing, and use it to predict what provides value to the consumer.

Late last year, Amazon opened its first brick and mortar store, with the top seller surprisingly not found on a shelf, but in customer data swirling around every purchase.  According to futurist and Forbes contributor Rob Salkowitz, books in the new Amazon store have a shelf tag which includes a website review, a star rating, and a barcode, but have no price listed.

Using a smartphone camera to scan the barcode, the Amazon app sends full reviews, specs, and pricing. At the same time Amazon instantaneously collects shopper records and their physical location.  It now has shopper preferences, buying history, status as a Prime member, credit card information and more and on that basis it can make recommendations, offer coupons, and is ready to close virtually any deal, all while the book is still in your hand. The opportunity to increase in-store sales using personalized data-driven upselling, could help make up for the billions of dollars the company spends on shipping costs.

Brad Power, partner and process innovation researcher with FCB (Foote, Cone and Belding) sees the need for re-engineering the customer engagement processes. “Consumer brands and retailers are analyzing the many sources of consumer data, then provide shoppers with continuous and personalized information and value to deepen the relationship. This is only possible with sophisticated algorithms and rules in expert systems and automation.” This is why there’s a need for Deep Learning.

Deep learning is an approach to machine learning that gets computers to learn from data – a kind of “unsupervised” learning approach. The enormous amounts of data are the real reason experts feel AI will succeed this time.  Approximately four exabytes of unstructured data is generated every day (tweets, emails, video, “human forms of expression”). That didn’t exist before and can now be used for deep learning.

This approach organizes layers of neural networks, analyzing enormous data sets, not just for simple statistics, but is a quest for patterns within the data. The aim is for these networks to learn to perform tasks better, faster, and on a larger scale than humans could imagine.

At MIT’s artificial intelligence conference in Boston last month, keynote speaker, Ray Kurzweil, the “ultimate thinking machine” and director of engineering at Google, said before a packed room that the biggest advancement in AI had been in deep learning. He explained that neural networks’ basic unit is a neuron, akin to a human brain which houses around 86 billion neurons. Modules are organized in a hierarchy and can learn and remember patterns.

According to Kurzweil, “Just a few years ago AI couldn’t even tell the difference between a dog and a cat. Now it can. Google and Microsoft’s image recognition programs can now look at complicated images and tell us that it’s two cats playing with a ball of yarn on top of a television set. This has led to great breakthroughs in AI.”

Image recognition played an important role when Google DeepMind program, AlphaGo beat European reining Go champ last October. Go, an ancient Chinese game was attractive to AI scientists because of its difficulty in evaluating board positions and moves (there are more possible board configurations than the number of atoms in the universe). To unseat Fan Hui, AlphaGo used image recognition to ‘perceive’ the board and combined Monte Carlo tree search with deep neural networks.  These networks were trained by human supervised learning and used reinforcement techniques. Once AlphaGo was trained, it then matched the system against itself.  It achieved a 99.8 winning rate against other Go programs.

Defeating Hui was the first time a computer program had defeated a human professional player in the game of Go – a feat previously thought to be at least a decade away.  Now, AlphaGo is facing its biggest challenge yet.  Starting Wednesday, March 9, AlphaGo is playing the top Go player in the world. The AI program will play Lee Sedol in a five game challenge match to be held from Wednesday, March 9 to Tuesday, March 15 in Seoul, South Korea. We have just heard that Sedol has just resigned in the first of these 5 matches, and that AlphaGo is now one nil up in the series. Perhaps like Kasparov before him, he may recover to win the series, but that would imply that within a year, it will be an AI machine, not a man, that is the best Go player in the world.

Online services like Facebook and Microsoft already use deep learning in their image recognition programs as well as voice recognition, and understanding natural language. Google recently released an AI toolkit called Google Cloud Vision API so developers can create apps that tap into the emotional content in a photo.  Since the announcement, thousands of companies have used the API, generating millions of requests for image annotations, Google said.

Rob High, CTO IBM Watson and also a keynote at Boston’s AI conference said there’s been a lot of advances in vision science surrounding AI. “Recognizing what’s in an image – stills as well as video. Being able to decompose an image and being able to recognize an event – what’s going on in the picture?” He said there’s been a rapid shift from human engineered features to generated features derived through deep learning. “Significant natural language classification, vision, speech recognition, language translation, are all now fundamentally based on variations of deep learning – neural networks technologies.”

High said that while a lot of the science around AI, academic and some commercial work, focuses on classic AI – computing systems that reason like humans.  “There’s another way of thinking about the role of AI – the role of intelligent augmentation.  How do we affect human beings on a psychological level to amplify our own human cognition?” This is the what Watson is now engaged in.

In 2011, after Watson’s historic Jeopardy win (won by reading more than 200 million pages from Wikipedia and other sources), and following some enhancements to the system, IBM realized there was too much demand for the kinds of applications people wanted to apply the Watson technology to so they opened the Watson Ecosystem – a way to provide businesses and entrepreneurs Watson’s cognitive technology. Google also announced its own machine learning software – TensorFlow, giving access to independent developers to build their own software, based on Google AI technology.

According to The Conversation (a not-for-profit media outlet which uses content from the research community) this open-source AI trend keeps tech companies on the cutting edge and paves the way to the future.

Kurzweil predicts fashion will be another industry that takes advantage of open-source technology using 3D printing. “We’ll get there by 2020.  You’ll be able to print out clothing – there will be many free open-source designs you can print for free, then you can print out high-quality clothing at pennies per pound.”

He points out there are many industries that have been transformed from physical products to digital products like books, movies, and music – and they all thrive in an open-source market.  “The co-existence of a free open-source market, which is a great leveler, providing high quality products for free and a proprietary market – that’s going to be a model for the economy going forward. The revenues in these industries have gone up not down.” But typically the revenue for a single product has come down dramatically.

Kurzweil noted 3D printing is going to revolutionize manufacturing going into 2020.  “You can already print out organs – lungs, hearts, and kidneys, using biodegradable scaffolding then populating them with stem cells and then grow out the organs.  Human testing will begin soon.”  Kurzweil is well known for this thesis on the future of humans being improved by computers, something he has espoused in a handful of popular books including The Age of Intelligent Machines. No-one doubts his knowledge or enthusiasm for the topic, but we have to doubt his forecasting ability since he wrote that book in 1990 and many things in it still have not yet come to pass. It might not therefore be in our lifetimes before this process is commonplace – but it is coming.

Not to be outdone by Google, IBM has also been developing AI systems around medicine. IBM’s CTO pointed out another offshoot of Watson’s Jeopardy win was the publicity and an excitement in the healthcare field.  Doctors recognized Watson was capable of fixing an ailment they had long tried to cure – keeping up and organizing medical literature. IBM’s Oncology Offerings was soon created.  It was a different cognitive system based on individual patients. “We could read a patient’s history and then look for similarities in other patient’s records”. They could then compare this with more than 10,000 variables that define the makeup of a human being – genetics, environment, disease history, etc.

The Watson team then met with top medical professionals and created a new cognitive system that could evaluate clinical expertise that would decipher the logic doctors apply when diagnosing a patient. High said, “Most doctors, are good at what they do, not because they’ve been well educated, but because they’ve been well experienced.”  High said the trick was pulling out 25 years of experience – which was now buried deeply in doctor’s subconscious – then applying that to the system.

AI is advancing so fast; we cannot predict what the world will look like five, ten or 20 years from now. What we can say is that no matter what the industry, AI will touch and transform it sometime soon. A collective view from MIT’s AI conference is that AI is coming full-force and every industry needs to prepare. “By the time we get to the 2040s, we’ll be able to multiply human intelligence a billion-fold. That will be a profound change that’s singular in nature. Computers are going to keep getting smaller and smaller.” Kurzweil predicts that “ultimately, they will go inside our bodies and brains and make us healthier, make us smarter.” Many readers will not be re-assured by that thought.

by Barbara Qualmann, ReThink Research