What Is Deep Learning? A Primer with Bonsai’s Mark Hammond

Over the past few years, the tech conversation has been dominated by references to AI and deep learning. Some people liken it to websites in 1995—everybody knew they needed one, but they didn’t all know what they were. Today, there are a lot of people talking about AI without having a real grasp on how deep learning works or even what it means. So what is it, really?

One company that has already begun using deep learning is Bonsai. Mark Hammond, the co-founder and CEO, had been studying AI since his time at Caltech, but realized that for machines to learn, they had to be taught. He wanted to build a tool that allowed non-AI experts to put this technology to use in real-world equipment—think giant manufacturing equipment or power grids. Bonsai’s AI platform lets experts build intelligent control capabilities or optimizations for their processes. As a company at the forefront of applying deep learning to industrial processes, what does Bonsai think it means for the world?

“AI has had a long journey, and generally speaking, the definition of AI has changed over time,” Hammond says. “Over the many decades, the funny joke in the industry is AI is whatever humans can do that computer still can’t. It’s a shifting target, constantly moving, but when you dig down into it, AI is a superset of machine learning, and machine learning is a superset of deep learning. AI is about enabling a system to have intelligent behavior, solve problems, have cognition, and think about things in reasonable ways.”

The insight that brought Bonsai into being was a fairly simple one. Hammond had been working for a long time trying to build the best possible learning algorithm, like so many others in the field. He began asking himself the question, “What do we do when we’ve got it? I want to do real-world work with this. What do I do?” The insight he had was to imagine being in a place to enable developers and subject matter experts everywhere to break down these problems and codify what they want to teach—not how it should be learned. He realized that if he could codify that, he could “make it possible for people to take that particular expertise and knowledge they have about whatever problems that may be going after, and enable them to focus on their expertise. It’s not focused on statistics and lots of machine-learning stuff. It’s focused on, what is your expertise? How do I enable you? It’s the abstraction that unlocks AI for everybody.”

Like the internet in 1995, AI is on the precipice of mass adoption. “The trick with AI, which is hard for a lot of AI people to accept but it’s true, is it’s just a technology,” says Hammond. “It’s a technology like any other. And it’s a really sophisticated, exciting, fun, sexy technology, but it’s a technology. The real trick is to focus on the problems people are trying to solve. The technologies are used to solve problems. Technology isn’t the problem. So, you don’t go to the customer and say, ‘Let me tell you all about this cool AI thing.’ You go to the customer, and you say, ‘What are you doing? What are the problems that you’re facing that have proven to be intractable? And, let’s see if AI now makes it possible to solve those problems.’’’

The best way to think about AI is as a new form of programming. “It’s a new way to add more tools to your arsenal that you can bring to bear when you’re trying to tackle problems that your business has been facing,” Hammond says. “Programming, the traditional way, is useful for systems or problems that are well understood and can be codified into repeatable steps. But for those problems that can’t be broken down and systematized so easily, AI technologies can now learn how a human tries to tackle them and help you codify what that person is doing effectively.”

So far, so good. But what is deep learning and how does it fit into this picture? For Hammon, “deep learning is a specific algorithmic technique used within machine learning. They form this hierarchy of pieces that you use. Machine learning can be broken down into lots of different techniques for the ways that you can learn.”

The most common technique used today is called supervised learning. “Supervised learning means that I have a set of data, and it has a set of labels associated with that. I know that this is a picture of a cat because you told me and I have a lot of them. Now, can you figure out how to map between pictures and cats so I can make those predictions in the future?”

Reinforcement learning is a different technique that is more commonly used in industrial applications of AI. It handles situations where you don’t necessarily know what the mapping should be—that is, you may not know how to get from the inputs to the desired output, or the desired output is in an interactive environment. “Think of playing a game like Go or chess,” says Hammond. “Over the sequence of many steps, I’m going to get to an outcome that hopefully is the one I wanted. But you won’t know whether any step you took was correct. Instead, you get to the end, and you say, ‘Well, how did I do? Did I get a good score or not?’ And you need to be able to work backward to learn, ‘Okay, what should I have done in between? How good was the policy I was following?’”

In industrial settings, reinforcement learning might be used in a system that you want to control or a process that you want to optimize. “At any given point in time, if you have a piece of manufacturing equipment or you have a robotic apparatus, telling the system ‘this is the exact value that the torque on this motor should be set to’ is not something even the expert human operators do,” says Hammond. What reinforcement learning–enabled machines do is they evaluate whether they have navigated to the desired outcome. This isn’t about decomposing the process down to the motor torque on a particular joint. “You don’t do that. Nobody does that,” says Hammond. “And so there has to be a way for you to say, ‘Did you get to the outcome I wanted or not, and did you do it in a way that is appropriate and reasonable or not?’”

To some extent, this seems like simply teaching machines to do our current jobs. But does it change anything about the nature of work as we know it? “It definitely shifts how we go about doing things because now things that previously were completely resistant to automation, we can automate,” says Hammond. “And there’s a lot of really tedious work that people don’t like doing or that’s impossible for people to do, but you still need to have a human operator sitting there operating this piece of equipment to get it done.”

“Modern reinforcement learning and deep reinforcement learning technology lets you capture expertise that people can’t really express so easily in a way that it can be learned by a machine,” Hammond goes on. “Your expert machinist or your chemical engineer or whoever the subject matter expert happens to be can break the problem down and tell the AI, ‘When you think about tackling this, you really need to think about it in terms of these concepts. And if I wanna teach you about these concepts, this is how we can do it.’ That enables so much capability that was previously not possible. You can now optimize things that have been resistant. You can carry out robotic tasks that were intractable before. Things that require lots of manual labor, you can now automate more readily. So it opens up the world of possibilities for a lot of these industrial applications.”