BACKGROUND IMAGE: agsandrew/stock.adobe.com
Generative adversarial networks, reinforcement learning and transfer learning are approaches that have been explored by theoreticians and researchers for years. Today, with recent improvements in technology, these new deep learning techniques are finally becoming practical for enterprise use.
"They are not really new concepts," said Hermann Ney, professor of computer science at Germany's RWTH Aachen University and director of science at speech recognition company AppTek. "But now, in the era of deep learning, they have a better chance to be helpful."
GANs blur line between real and artificial
Last year, researchers from chipmaker Nvidia, based in Santa Clara, Calif., released a video showing computer-generated faces, cars and furniture suites that were amazingly realistic.
The secret? Generative adversarial networks (GANs), in which two different AI systems battle it out. One system tries to create realistic-looking images; the other system tries to tell which ones are fake and which ones are real.
As the two duke it out, both get better and better, and the results can be exceptionally lifelike -- and a little disturbing, said Vivek Katyal, global analytics leader for risk and financial advisory at Deloitte.
"It's a pretty scary thing" how realistic these artificially created images can be, he said.
But while GANs are of clear benefit to, say, film and video game companies looking to fill out crowd scenes, there are other uses as well.
For example, companies can use it to take photographs and create 3D renderings or even generate models for 3D printing.
"This is now being looked at in very advanced manufacturing," Katyal said. There are also potential uses in other areas, he added, such as medical imagery and generating the massive sets of training data that drive deep learning. "The key application is generating image sets for learning data."
However, Katyal warned that companies need to be wary of inadvertently introducing bias and errors into their systems. If an enterprise implements GANs without first ensuring their data is clean, representative and unbiased, the deep learning technique could magnify these problems.
"I don't see it inhibiting adoption," Katyal said. "But that's because I don't think people today look at risk first. They look at what they can get out of it."
Reinforcement learning creates new strategies
Last month, Google published the results of its experience with AlphaZero, a system that learned to play Go and chess all by itself, without studying human games or getting any feedback from people.
The secret was its use of reinforcement learning, one of the most cutting-edge new deep learning techniques. The program played the games over and over in an attempt to beat its own previous versions. It quickly evolved into a system that could beat all existing competitors.
David Silver, who leads the reinforcement learning group for Google's DeepMind, wrote in a recent blog post that this approach can lead to a more creative kind of algorithmic game playing.
Reinforcement learning can be used by an AI system to teach itself how to do almost anything, as long as there's a way to keep score.
Practical applications include navigation software that can enable robots to find their way around new places that they don't have much data on yet or manufacturing robots that learn how to interact with objects, said Jacob Perkins, CTO at San Francisco security analytics company Insight Engines and author of several books about machine learning software development.
"I don't know if Roomba uses reinforcement learning, but it would be a good application of it," he said.
In fact, any process that can be optimized could be a target for reinforcement learning, said Christian Shelton, professor of computer science at University of California, Riverside.
Optimization challenges, such as supply chain management, data center energy use and cloud workload schedules, are currently being handled by other approaches, such as traditional statistical methods that don't require reinforcement learning. However, as these challenges get more complex because, say, companies look at more contributing factors, reinforcement learning will start coming into its own, Shelton said.
Transfer learning could lead to more natural AI
Transfer learning is something that comes naturally to humans. Once we learn how to do one thing, we have an easier time learning a second related thing, instead of having to learn each element of the second task from scratch.
This doesn't come naturally to computer programs, but AI programmers are using various methods to give new systems a head start.
With the recent advancements of new deep learning techniques, the possibilities of transferring knowledge have gotten better.
AppTek, for example, is a Virginia-based company that uses AI systems to understand and translate spoken language.
Transfer learning enables it to train its systems on large, publicly available data sets, such as broadcast and entertainment videos and audio. Then, the learning can be applied to other situations, such as user-generated videos or telephone calls, where the sound quality is different, said AppTek CEO Mudar Yaghi.
"The results in all cases are more accurate predictions -- such as chatbots that now recognize regional dialects or spellings," said Ken Sanford, analytics architect and sales engineering lead at New York data platform company Dataiku and professor at Boston College.
Sanford has worked with many companies on AI projects, including Walmart, in his work for AI vendors and as an independent expert.
Walmart and other retailers are using transfer learning to help better categorize products, he said. "They have too many new products, and the classification system is too complex to do it manually." Transfer learning, in combination with image recognition, can identify subtle differences among products.
Cloud providers that offer AI model-building services, including Google, Microsoft and Amazon, are also using transfer learning.
"You upload a training set, they reference the history of all models they have that appear close to it and they use the labeled training set to further hone the model," Sanford said.
New deep learning techniques may lead to more natural AI
Over the past decades, AI technology has improved dramatically, moving from basic rules-based systems to statistical approaches. More recently, it's come to be powered by the machine learning algorithms widely in use today, Deloitte's Katyal said.
Now, GANs, reinforcement learning and transfer learning are helping take us beyond machine learning and into narrow AI, he said. That's the next step on the road to general AI.
"General AI is when it almost automates human intelligence," Katyal said.