Listen to this story
“Where Is the Future?” is a series of interviews with industry leaders considering the potential and complexity of technology on the horizon.
Two summers ago, Courtenay Cotton led a workshop on machine learning that I attended with a New York–based group called the Women and Surveillance Initiative. It was a welcome introduction to the subject and a rare opportunity to cut through the hype to understand both the value of machine learning and the complications of this field of research. In our recent interview, Cotton, who now works as lead data scientist at n-Join, once again offered her clear thinking on machine learning and where it is headed.
What kind of problems is machine learning designed to solve? And are there times when machine learning isn’t the right method for prediction?
Courtenay Cotton: Machine learning is used for prediction. It could also be used for different types of problems. I think the main criteria about what type of problem you can easily solve with it has to do with what kind of data you have around that problem. Any problem where there’s really clear output that you want, that you know what it is, it’s well-defined, and you feel like you have enough inputs of whatever sort to help you learn it is a problem that’s really well suited to machine learning.
The other methods of research you might call by different names. You might say, “I’m using statistics and doing the regression,” but to my mind, that falls under the heading of machine learning as it is used broadly today. It’s not like machine learning is this new thing. It’s mathematical methods to derive relationships between input and output data that can then be learned from data. It’s generally the same as statistical methods. There’s not this really sharp division between new-style — what people call machine learning — versus mathematical methods that have been around forever, that have been accelerated by current computer technology, obviously, but are still fundamentally the same kind of methods. People develop new algorithms and have breakthroughs, but it’s always that you’re optimizing algorithms, you’re solving for functions. It’s not a different kind of thing.
The power there, the reason it’s new and shiny and a hot topic, is just that now you can have computers crunch giant data sets and learn what seems like more magical kind of outputs to humans because of the power of being able to learn simultaneously from all this data.
Will Google and Facebook always have an advantage due to the size of their data sets, or is a more specialized data set also relevant here?
They definitely have large advantages. It may also depend on the problem you’re tackling. I can imagine collecting a bunch of data in a specific domain that people aren’t looking at right now and having an advantage in that domain. If you’re trying to do a lot of standard, traditional things that people are interested in doing right now with a kind of internet user data, the guys with all the internet users are gonna win that game, at least for the time being. But there will always be a new company that will be a threat to these guys. It’s the cycle of companies becoming dominant in some field, and then getting challenged by newer companies.
Could you talk about issues of classifying data and coming across errors?
Data cleaning and data wrangling, as the first step doing any of this stuff, is a giant part of this field. There’s almost never not errors in your data. It would be really hard to not have any misleading things. Sometimes there are systematic errors or systematic biases; [for example] all your users were from one demographic, and now you’ve learned something, but it’s not applicable to other demographics. Just knowing what you’re going for and looking at your data in a principled way and seeing if it is going to be able to predict that without flaws and biases is obviously a big problem.
In the tech community about 10 years ago, there was a cliché — not always true — that everyone was a college dropout. But it seems like machine learning is really driven by academics.
It is largely an academic field. The breakthroughs in it are usually very serious math optimization kinds of stuff. Which is not to say there aren’t self-taught people, but it’s really more academic than the recent tech booms in general. It is also getting democratized fast. There are so many new libraries and tools and plug-and-play, out-of-the-box “we’ll run this tool.” It leads to the problem of people using algorithms that they don’t really understand what’s going on underneath. They’re just like, “Oh, I heard this one works. Let’s try that one,” and not really understanding, which could be okay or could lead to more problems like we were just discussing.
I come from an academic background. I have a PhD, although I’m not in academia anymore. There’s a certain flavor to academia that, I guess, is driving the really hardcore machine-learning community. Actually, there’s a movement where a lot of the front-tier machine-learning research is leaving academia and going to the big companies. Facebook and Google are stealing a lot of academics to do this work. Maybe that culture comes over to industry. There’s a lot of cross-pollination.
What are some common misperceptions about machine learning?
There’s always an air of mystery because, in reality, even for us researchers, a lot of these algorithms are black boxes. There’s not a lot of insight into what’s going on in them. It’s not at all confusing that people get confused about what’s happening there, because sometimes it’s really not obvious. The biggest popular misconception, I think, is “these things are going to overtake humans any moment now,” which is not realistic at all. Yes, they will overtake us in specific domains, but it’s not anything close to general intelligence, general cognition. Like most other technology, there’s also a general misconception of “oh, it’s magic,” but that’s just because it’s generally hard to grasp how these things work.
By specific domains, do you mean the context?
IBM hypes [that] its Watson can win Jeopardy, but if I want Watson to come have a dinner conversation — is that going to go as well? I can train a machine to beat me at chess or at driving my car down the street, but that same machine is going to be inept at the other thing, and it’s also going to be inept at 10 million other things that humans can do easily — that babies can do.
Some AI researchers are legitimately trying to figure out how you would get a machine that learned like a human child. But in general, most of the work is “I need this very specific thing that just does this one thing, and I’m going to throw all the data in the world that I can get my hands on at it.” At the end, it will be pretty good at that one thing—if we have the right data. But it’s not conceiving of the problem the way a human would, it’s not really solving the problem in a way a human would, and there’s just not a lot of direct correlation there.
About this Collection