Machine learning: researchers and developers rave about it, but the average person may not use it or even know what it does. At Google in Cambridge, Fernanda Viégas SM '00, PhD '05 is changing that.
“One of my research goals has always been to democratize information tools, especially those deemed too complex for lay people,” she says. In 2017, Viégas became co-leader of Google’s People + AI Research initiative. Their goal was simple: help make machine learning (ML) accessible and easy to use. That year, they created deeplearn.js.
“When we started working on a JavaScript library for ML, people thought we were crazy,” she explains. “ML is usually done in Python and runs on servers. And yet, here we were, bringing the technology to the web and having it run locally on the browser.”
This analytical framework of deeplearn.js runs pre-trained models or allows the user to train neural networks to analyze and “learn” from data. When the framework is combined with access to large amounts of raw information, it can make classification, categorization, and prediction on its own.
For example, Teachable Machine, created by some of the team behind deeplearn.js, allows a user to train a computer to respond to facial, hand, and body movements. After recording movement on the computer camera, the system "learns" to differentiate between motions. The user assigns responses to each movement: raising your hand might show an image, display a sound, or play a gif.
I can’t wait to see where the students will take the technology, says Viégas.
In essence, deeplearn.js has helped lower the bar for ML experimentation. So far, the results have been impressive. One developer in Portland, Oregon, developed an accessibility browser extension that allows users to control the cursor with facial movements. Those with limited mobility, like stroke sufferers, can use the extension to surf the web.
One of Viégas’s collaborators, Hal Abelson PhD '73, teaches an undergraduate class based on this technology called 6.S198 Deep Learning Practicum. NYU is using the tool to develop their own high-level JavaScript library, called ml5.js, for creative student coders.
“I can’t wait to see where the students will take the technology,” says Viégas.
Last week, deeplearn.js became part of open-source machine learning framework TensorFlow.js. This will allow for better functionality, including integration with other TensorFlow models and robust technical support.
This work integrates with Viégas’s larger focus as leader of Google’s Big Picture team, part of the deep learning AI group Google Brain. She says, “There’s a growing number of ML researchers and teams in the Google Cambridge office. Big Picture certainly brings its own special flavor to the mix: with a focus on human/AI interaction and data visualization, our team is in a great position to help shape the future of AI.”
Viégas says Kendall Square is the perfect place to do this work. “We are huge believers in the power of Cambridge and Boston to serve as technological hubs. The area is unique is its academic + industry makeup. This is why the Google office in Cambridge continues to expand and attract top talent.”
She’s been a part of the Kendall community since her time at the Media Lab. Originally an artist, Viégas says MIT helped her combine her visual expertise with a better understanding of data and computational analysis. She says, “I felt empowered to combine my graphic design background with my newly acquired computational skills to work on data visualization.”
