The impact deep learning is having on artificial intelligence
At the VB Transform conference in San Francisco, MissingLink.ai founder Yosi demonstrated how easy it is for machines to learn skills that just a few years ago, only people could pick up. If you’re a business, he added, you’d better start investing significantly in AI right now, or another company will outcompete you.
Competing with AI
Yosi, who served for almost 10 years in the Israeli Defense Forces as a software developer, co-founded MissingLink.ai in 2016 to help software engineers train artificial intelligence to do their jobs faster. Since then, MissingLink has joined the Samsung NEXT product team and launched its platform to help data scientists manage their deep learning operations — what it calls “DeepOps.”
Getting a deep learning project off the ground, Yosi told the audience at VB Transform, is relatively easy. He said it is possible to get started in less than a month thanks to deep learning tools developed in the past few years that enable software to learn by example — thousands of examples at a time — rather than having to be taught.
Deep learning, a form of artificial intelligence, uses artificial neural networks that mimic the function of the human brain. Just as the brain works by forming electrochemical connections between millions of brain cells, called neurons, deep learning works by forming associations between nodes defined by software.
Now, teaching an AI machine new tricks — such as recognizing objects or how to play a game — is as simple as showing it enough examples or demonstrating various permutations over and over. But it wasn’t always so easy.
Until recently, software engineers had to hand-train artificial intelligence systems to recognize objects such as cats. “We extracted features, all those features that we think are part of a cat,” he said. “And we fed it into statistical models that said, you know, what, 60%, this is a cat.”
But that time-consuming process became obsolete in 2012, said Yosi. That’s when deep learning became mainstream — thanks in part to the work of Stanford University computer scientist Fei-Fei Li, who led the creation of a dataset comprised of 15 million images grouped into 22,000 categories.
With a common dataset to train their machines on, researchers everywhere could advance the field at a breakneck pace. By 2017, Yosi said, artificial intelligence could recognize images with an accuracy of 2.25 percent. “Now, you need to understand that as human beings,” he noted, “our error rate is 5 percent.”
Garbage in, intelligence out
The reason deep learning is so successful is that it requires no hand-holding. “Unlike machine learning, we’re not trying to make sense of what’s inside,” said Yosi. “We’re just feeding it raw images, we’re not writing any code.”
Deep learning algorithms can do their thing with just a few tens of lines of code. They learn by trial and error, getting more and more accurate as they go, moving at machine speed.
Take a deep learning system’s guess at whether an image is a cat or not. Say on the first attempt, it guesses with 20 percent accuracy—far worse than the hit rate of a typical human. On the next try, however, it will inch closer to getting it right.
“Now, if I ask again about the same image,” said Yosi, “it won’t be 20 percent. It might be 21 percent, or a bit less than that. But it’s going towards that right result. And we’ll do it again. And again. And again. And again. And again. That’s training.”
As an example of just how far an AI can get in learning to perform tasks with superhuman accuracy, Yosi pointed to AlphaGo, a program developed by DeepMind Technologies, a UK-based company owned by Alphabet, Google’s parent company.
AlphaGo started out with no knowledge of how to play the ancient Chinese game, Go. With 10 to the power of 170 possible configurations on a board with 64 squares, Go has far more possible configurations than chess, making it an ideal proving ground for AI.
After exposing the program to 160,000 amateur games to give it the idea, AlphaGo’s developers then had the computer play itself, over and over and over. By the end of the process, in 2015, the machine trounced the European Go champion in a 5-0 match, and then went on the defeat the world champion 4-1.
“After three days of training against itself,” said Yosi, “it became the best in the world.” Now, he added, “There’s no game that a machine cannot play better than a human being.”
Today, deep learning has moved on to other conquests. As Yosi demonstrated during his presentation at VB Transform, it continues to get easier to train a deep learning AI. To underscore he point, he quickly used an AI-powered Google website called Teachable Machine to create a bot that could recognize hand signals.
Thanks to AI and deep learning, data is the new currency of successful companies. “This is the end of code,” Yosi said. “It’s just about being able to get your hands on data. This is going to change everything for everyone.”
Anyone in charge of a company, Yosi said, had better invest at least 15 percent of his or her budget in making sense of data. “Your data is your future success, period,” he said. “You will not be able to compete with another company that does exactly the same thing as you do using AI.”
Learn more about the future of AI by watching the End of the Beginning video series.