American Scientist

Jeff Dean, head of artificial intelligence at Google, gave the keynote address at Sigma Xi’s 2018 meeting, “Big Data and the Future of Research.” Dean, who joined Google in 1999, completed a PhD in computing science at the University of Washington in 1996, and he has worked on a number of prominent projects in large-scale data processing and machine learning. At the conference he spoke to American Scientist’s editor-in-chief, Fenella Saunders, about some of the major advances, and concerns, facing current artificial intelligence research, and how it interfaces with human society.

The field of machine learning has made a lot of progress. Where do you see it going now?

We’ve seen significant developments in deep learning—essentially, a rebranding of artificial neural networks. These have been around for 30 or 40 years as a way of describing abstract ways of learning from interesting inputs and outputs. But now it turns out that deep learning is useful for all kinds of problems in the fields of computer vision, speech recognition, language understanding, and language translation.

The fact that computer vision and these language-related tasks are becoming more successful means that computers can now perceive the world around them much better than they could before. And that has implications beyond the boundaries of computer science.

Now, all of a sudden, we can tackle many of the grand challenges in the fields of design and engineering by means of machine learning. Questions like “How can we make health care better for people?” or “How can we develop or rework urban infrastructure?” are amenable to machine learning. As an example, autonomous vehicles are going to be a big factor in rethinking how cities should be designed, because they’re going to be very different from cars that we drive.

Machine learning was built on the model of showing a computer lots of examples until it figures out the connections. Are there other mechanisms now?

Most successful kinds of machine learning are of that form: Collect a large data set of inputs and outputs that you care about. It might be a bunch of pictures, and each picture is labeled with “that’s a truck,” “that’s a pigeon,” “that’s a particular kind of monkey.” Through exposure to many examples like that, the system can learn to generalize to a completely new picture. Now it can say, “Well, that picture is also a truck.” That process is called supervised machine learning.

But other kinds of machine learning are also making significant strides. In a technique called reinforcement learning, you have a set of actions you can take, and at each step you try to predict what actions make the most sense. As you proceed, you begin to get a sense of whether the set of actions you took was a good idea or not.

If you imagine playing the game Go, eventually you win or lose, and then you can attribute credit to all the decisions you made in playing the game. That model can be used to teach systems to learn on their own: If you pit a couple of computers against each other, they can both learn the game by playing it together.

How can data sets and algorithms be made less prone to unconscious bias?

Bias in machine learning algorithms is a big problem. It’s something that we need to be aware of, and there’s an active area of research about how we can make these systems in automated ways be less biased. One of the issues is that sometimes the data you’re training on are biased. These are data from the real world—the world as it is, rather than the world that we wish it were.

Or you may have collected a data set that doesn’t match the distribution of data that you trained the model on. For example, if you trained the model on photos of North American weddings, and then started trying to recognize Indian weddings, the distribution of the data sets and the imagery you might see are quite different. So collecting data sets that have the right distributional properties for the problem you care about is an important skill that you should apply for machine learning.

There are also techniques where you can algorithmically adjust machine learning models in such a way that, for example, all other things being equal, you’d like these two groups of examples to have an equal chance of obtaining a certain kind of outcome. That is a way to take a model that has been trained in a biased way and adjust the output of it algorithmically to make it less biased.

What is your take on privacy in the age of machine learning?

As people take advantage of the many new online services out there, they create data about how those services are being used. And often companies use those data to improve the services, such as by understanding that when people watch this kind of video, they may also be interested in that kind of video. Or when they mistype this word and then later correct it, maybe we can learn a spelling correction system that helps all users based on the behavior of how people do spelling correction.

I believe people should have control over whether those data are collected, and if they are, people should have the ability to delete them. And that deletion should take effect in a timely manner as people update their machine learning systems. It is true that there are lots and lots of online systems, and people now grow up in an online environment where they express themselves on social media and so on. The behavior and decorum around privacy have changed over the past 10 or 20 years.

Are there ways machine learning can help people break out of silos?

Many products have more of a focus on recommendations than on breadth of sampling. It’s very easy for machine learning to operate so that if you like this sort of thing, we just recommend more of that thing. I think there is an opportunity in the algorithms to encourage diversity of the things that people are exposed to. And it’s definitely something that we think about at the product level. We want to expose people to information they might find thought-provoking, as opposed to things that completely agree with their current thinking.

Even though machines are now training themselves, the machine learning programs are still created by human beings. Do you place an emphasis on having a diversity of people from various backgrounds and cultures on your team to ensure that you produce more robust research?

Yes, absolutely. Computer science and machine learning are creating really interesting new products and applications in the world, and they’re affecting billions of people. And so you want those services to be created by the breadth of people that use them. It’s important to me that we find talented people all over the world with all kinds of different backgrounds to help create new kinds of machine learning algorithms, new kinds of approaches, new kinds of products.

As an example of that, in June 2018 we announced our first research lab in Africa, in Accra, Ghana. I’m really excited to have that. The field of machine learning in the African continent has been growing very quickly. There’s a lot of interest there, and a great many researchers, young students, and young practitioners. I went to Cape Town, South Africa, to participate in Deep Learning Indaba 2018, with about 500 people from 40-ish countries around Africa all coming together to hear a series of lectures. It was a fantastic experience. I want that diversity of people and backgrounds to all contribute to the field of machine learning and computing.

An extended interview: