Googler Fernanda Viégas Designs Human-Centered AI for Everyone

Fernanda Viégas is a co-founder of Google’s PAIR (People+AI Research) initiative, which is part of Google AI. Her work in machine learning focuses on improving the interaction between people and AI — with a broader agenda of democratizing AI technology. She is known for her contributions to social and collaborative visualization, and the systems she and her team have created are used daily by millions of people. Fernanda’s passion for making complex data understandable to everyone has led her to visualize wind currents, study collaboration patterns in Wikipedia, and create a dynamic map of news around the world. Her visualization-based artwork with long-time colleague Martin Wattenberg is part of the permanent collection at the Museum of Modern Art in New York and has also appeared in exhibitions around the world. Fernanda, who was born in Brazil, holds a PhD from the MIT Media Lab.

How and why did you begin working at Google? 

I started at Google in 2010, and my path to Google was not traditional. 

I had just started a data visualization studio, self-funded, and literally in my living room with my colleague Martin Wattenberg. Martin and I had been working together for seven years, and the two of us had just come out of IBM Research. We were getting acquainted with life as a very tiny start up when the phone rang, and it was Google! We were so excited because we thought we had our first big client. But the Googlers on the phone said, “No, you don't understand...we want you to come work here!”

They explained why it made sense for us to join Google and do data visualization here. We’ve always been excited about public-facing data visualization, and making complex technology and information accessible. So, after some discussion, we were 'acquired', and that’s how Martin and I ended up in research at Google.

Can you give an overview of your role and mission at Google?

At first, we worked on data visualization with many different Google teams. Then, about two and a half years ago, we joined the Google Brain team.

Although the initial focus of our work is using data visualization to help visualize machine learning, our impact is broader. I co-lead the PAIR initiative. This is a cross-Google initiative focused on building human-centered AI systems. These are machine learning systems that we design from the ground up, based upon users’ needs, to be productive and fair.

How does your background influence your work? 

A lot! My background is not typical for someone working in machine learning.

My background is in graphic design and art history. I did not program until the end of my senior undergraduate year at the University of Kansas.

I trained as a traditional graphic designer - doing things like silk screen printing, poster design and book design. I loved all of those things, but before I graduated, I heard about this place called the MIT Media Lab.

Students come to the Media Lab from many different backgrounds (like musicians, educators and biologists). They combine their domain knowledge with technology, to push the envelope of what is possible. This was very interesting to me - except that I had no idea how to program, so it required a very steep learning curve to take me from my traditional graphic design background to MIT. That is how my immersion in the tech world started.

My work at Google is highly interdisciplinary, and I bring a fresh perspective to the problems we're trying to solve. The fact that I have a graphic design background, the fact that I am an expert in data visualization, and the fact that I know HCI (Human Computer Interaction) - these all have an impact on my machine learning research.

As technologists, we can build better AIs, and better machine learning systems, if we start from the human need - if we start from the user.

Why is AI important?

AI is important in at least two different ways.

One is that AI allows things that were not possible before to become possible. Until now, everything we did computationally was rule-based. When we programmed a system, we first identified all that system’s possible states and rules. This is great if you know all the rules, but there are many things that we need to achieve where we don't know all the rules.

For instance, how do you recognize someone’s face? You can’t explain all the rules you use to recognize a person’s face. In the same way that we learn to recognize people, we also learn to tell the difference between a dog and a cat, but we don't know exactly which rules are involved. We just see dogs, and we see cats, and eventually we are able to tell them apart. That's exactly how AI systems work. We show them many examples, and they start to understand the likelihood that they are seeing a dog, or a cat. As they are learning, they're coming up with patterns,  the same way our brain does. AI systems learn from existing data, including historically-biased data, so it’s particularly crucial to train them with diverse and inclusive data sets.

At the same time, because AI works like our brains, it allows us to solve very complex problems that we could not solve previously.

Another reason AI is important to me is that I feel that we are just starting to scratch the surface of AI’s creative and expressive possibilities. If I am a painter, what kinds of paintings can I create? If I am a musician, what kind of music can I make? If I am an architect, can these machines help me create a more beautiful and functional house? As a designer or creator, can I be even more expressive than I am today?

These are only two aspects of AI’s amazing potential.

What inspires you most about your work?

The possibility of bringing all people, no matter their background, closer to this very powerful technology.

I'm interested in making this technology useful to people who need to accomplish important things every day. Doctors are trying to diagnose diseases; farmers are trying to grow healthier crops; people are traveling to different countries and need to speak different languages. Building this interface between people and technology inspires me.

Another thing that deeply inspires me, and this was true before I became involved with machine learning, is broadening the community that can engage with complex information. For instance, data visualization is very good at engaging people with statistics, without requiring that they have a PhD in statistics. Our eyes can quickly pick out a pattern such as a trend, a distribution, or an outlier. These statistical concepts can intimidate people. However, if we just speak the visual language of our eyes, non-technically trained people can intuitively understand these sophisticated concepts.

I am also currently researching how can we democratize the power and the possibilities of machine learning for a broader set of people.

Right now, using machine learning (ML) requires a steep learning curve. We are working on ways to make machine learning more accessible. For instance, my team just released TensorFlow.js. This is the first web-based machine learning library that runs in your browser or on your phone. It sends no data back to the server and opens up machine learning to a whole new set of people, a whole new set of developers. For example, Javascript developers are usually not familiar with machine learning because ML is usually done with a different language called Python, but TensorFlow.js creates that bridge and makes it happen easily in your browser. This is an example of making machine learning technology accessible.

How can people engage with and benefit from AI? 

People already are!

Very humble things that we do every day use AI, for example: spam filters. When you check your email, you can be sure that there is an AI system helping you in the background, sorting spam from non-spam.

There are also very tailored and customized ways that you can use AI to solve your own specific problems. 

Here is one example. In cucumber farming, sorting cucumbers by texture, color, shape and size is very important. The son of a Japanese cucumber farmer, seeing his mother spend eight hours a day sorting cucumbers by hand during harvest season, and inspired by AlphaGo’s competition with the world's best Go player, used TensorFlow to build a very simple image recognition system that sorts large, medium and small cucumbers by size.

Let me give you another very interesting and specific AI application. A developer used TensorFlow, and his laptop camera, to enable slight movements of his head to control his computer’s mouse. If he looked down, his cursor would move down. If he looked to the left, the cursor would move to the left. If he frowned, the mouse would click. Why did he do that? He was building this application for a friend who had suffered a stroke and had become quadriplegic. The only thing his friend could move was his head. And so, by building this TensorFlow-based tool, he enabled his friend to use his computer and the web.

I am personally inspired to make the power offered by artificial intelligence and machine learning accessible to everyone.

You can watch Fernanda moderate a panel at I/O 2018 : Opportunities, Challenges, and Strategies to Develop AI for Everyone, here: