From Deep Dream to Deep Drumpf, everyone's talking about neural networks these days. But what the heck are neural networks and what do they mean for the future of computing and design? Here's your quick primer.
What is a neural network?
This is a sticky issue, because there's no single definition neural network that's universally agreed upon. In essence, a neural network is a computer program that tries to simulate the way a human mind works—more specifically, by simulating neurons themselves.
What's a neuron?
In your brain, there are hundreds of billions of tiny cells called neurons, each of which is connected to maybe tens of thousands of its brethren in complicated, ever-changing webs. This charming interactive story is a great primer on how they work, but put (very) simply, neurons are how we learn. Each neuron represents a different idea, memory, or sensation. When two neurons fire at the same time, they link together, creating a mental association.
How does a neural network simulate our neurons?
It depends on the kind of neural network you're talking about, but let's take Google's Deep Dream for instance. Here, Google's engineers created a stack of artificial neurons (each of which might be a separate CPU, or a core on that CPU), arranged in layers.
These layers work together to figure out what the neural network might be "seeing" in any given image it's shown. Each layer of neurons makes its own set of increasingly specific inferences about the contents of that image, which the next layer builds upon to formulate a better "guess" at what the image shows. After Deep Dream was trained on a significantly large number of previously identified images, it eventually knew enough to recognize what a "banana" or a "parachute" looked like—and it could even draw one itself.
How is that different from the way computers usually work?
Again, this is very simplistic, but in a computer processor, the closest analog to neurons are transistors—which are hardwired together in straight lines. They're two-dimensional, not three-dimensional, and unlike human neurons, transistors never rewire themselves; each transistor is connected to just two or three others. That's compared to the 10,000 connections the average neuron in your own brain might have. The result? Computers are really good at thinking logically, moving from one transistor to the next in a straight line. But they'rereally bad at thinking creatively and learning new things, since they're hardwired for linear, logical "thought." Neural networks aim to solve that problem.
Can you give me a big, dumb analogy that makes sense of all of this?
Okay. The way a computer thinks is a little like riding a subway. To get downtown, a computer might need to pass through a dozen stations and change lines two or three times. The way a human thinks is more like teleporting. Once a human brain has made that "subway trip" downtown once—connecting two neurons on a mental map—it can just teleport directly there without making all the in-between stops.
Putting this back into literal terms, neurons can rewire themselves on the fly to work together more efficiently. That's how we learn, and that's how we combine new ideas to be creative. If we can simulate that same effect in a neural network, then, we can teach computers to think like we do.
What tasks would neural networks be better at than regular computers?
Pretty much anything that a human can currently do better than a computer—writing a story, analyzing the meaning of a work of art, even driving a car, or understanding a human being from context and body language. Again, Deep Dream is a good example. Google's neural network doesn't just see pixels when it looks at an image. It can actually tell you what the image contains. It's something humans take for granted, but up until recently, it was something computers couldn't accomplish.
What are some good examples of neural networks?
In addition to Deep Dream, Google's self-driving cars also use neural networks to detect pedestrians. But neural networks aren't exclusive to Google. Facebook uses them to discern faces and identify your friends. Neural networks have also been used to design new fonts, create digital graffiti, and even impersonate Donald Trump on Twitter.
How will neural networks evolve over the next few years?
Neural networks are the key to making computers more like humans, and automating the human brain's problem-solving and creative capabilities. Combine them with conversational interfaces, and neural networks can make true artificial intelligence finally possible—a revolution that would have a knock-on effect in the way we pretty much do everything. Designers in the future won't just use neural networks; neural networks may very well be designers themselves.