March 08, 2019
“AI will likely transform our society. We’ve already seen and experienced recent progress in the field, and we would like everyone to have a role in this transformation. Be a part of the change.” –Adriana Romero Soriano, research scientist at Facebook Montreal
– – –
Last year, Facebook AI research scientist Joelle Pineau authored an article on why diversity matters in AI. This year, on International Women’s Day, we turn the spotlight on Facebook AI researchers Natalia Neverova, Priya Goyal, Devi Parikh and Adriana Romero Soriano to, in Pineau’s words, “demystify what we do” at Facebook AI Research and offer career advice for other women hoping to enter the field.
This article outlines each of the researchers’ unique journeys, illustrating the value of internships, advocates and mentors, and also showcasing the culture and research environment at Facebook AI. The researchers also share their own perspectives on why diversity matters in AI, as well as the importance of awareness, representation, role models and community.
With many journeys, you don’t know at the start where you’ll eventually end up. Such was the case for Natalia Neverova, research scientist at Facebook Paris, who recalls initially being interested in math and physics in high school. Neverova shares that at the beginning of her undergraduate career, she didn’t think software engineering was for her, so she decided to pursue a more general engineering career—specifically in optical engineering.
After working in the analog world for a while, Neverova eventually decided she really enjoyed the analytical, algorithmic side of engineering, which is of course core to the field of software engineering. She switched to computer vision, and from there became interested in machine learning and AI in general.
Neverova obtained her master’s in computer vision and machine learning before moving on to a PhD in the same area. She then became a research scientist at Facebook AI Research, where she works on solving core visual perception tasks or reasoning tasks, such as object detection, segmentation, pose estimation and human-centered tasks in particular.
For those who have the means, Neverova recommends traveling as much as possible for work. Having completed her undergrad in Russia, her master’s in France and Norway, and her PhD in France and Canada before completing an internship in the U.S., Neverova mentions how traveling helped broaden her horizons, which can contribute to a diversity of thought.
On the topic of diversity in AI, Neverova explains how bias can influence advancement in the field. “Ensuring diversity is essential to handling our cognitive biases, of course on a personal level but also on the level of our research community,” she says. “The algorithms and the datasets we use to train our models reflect our own vision of the world, and this can be dramatically different for people coming from different backgrounds, from different social groups, geographical locations, scientific schools and so on.”
Growing up in India, Priya Goyal was encouraged by her mother to study and be independent. “I grew up in a very male-dominated society,” Goyal explains. “So my mother would always tell me I needed to go study so I could make a life for myself.” Under this guidance, Goyal found a passion for computer science and coding at the Indian Institute of Technology (IIT) in Kanpur. After graduating with a bachelor’s degree in mathematics and computing, Goyal then returned to IIT for her master’s.
Like many software engineers, Goyal’s journey into tech began with an internship. As she completed her master’s coursework, Goyal interned on the machine translation team at Facebook, where she contributed to a project focused on detecting profanity in Facebook’s machine translations. It was here where she discovered her passion for machine learning research and engineering. After graduating with her master’s in June 2015, Goyal joined another team that works on Pages and Places full-time, where she worked on fine-grained category prediction for Facebook’s large-scale places database.
Eventually, Goyal decided that she wanted more structure to what she was learning, so in 2016 she transitioned to Facebook AI Research in New York and focused on computer vision. Goyal has already made a significant impact in the field. One such example is the paper Accurate, Large Minibatch SGD: Training ImageNet in 1 Hour, which appeared at Data @Scale in 2017 and attracted substantial media attention. Goyal and her research team also won Best Student Paper Award at ICCV 2017 for their paper, Focal Loss for Dense Object Detection. Goyal also pioneered work in compilers for AI via their project Tensor Comprehensions, released publicly in 2018.
On the topic of building inclusive work environments, Goyal emphasizes the importance of creating awareness and understanding your own negative experiences. “I’ve been in places and positions where I was the only woman on a very large team, and I have strongly felt what it feels to be isolated,” Goyal shares. Looking back at her early career, she notes that her awareness of these issues has evolved. “Sometimes when Sheryl [Sandberg, Facebook’s COO] would comment on the difficulties women face in the workplace, [at first] I didn’t know what that meant and what she was talking about.” Being aware of these issues, Goyal says, helps in understanding that these negative experiences are oftentimes systemic. This knowledge can then be used to empower others and create a sense of responsibility within a community.
“This is a reality, so let’s take part in changing that,” says Goyal. “Let’s own it, let’s lift each other up – but at the same time create awareness.”
Raised in India, Devi Parikh took an interest in science and mathematics at a young age. She attended Rowan University where she studied electrical and computer engineering, and describes her foray into computer vision research as “one thing leading to another.”
Parikh recalls that as an undergrad at Rowan, she was approached by a professor who asked if she would be interested in participating in a machine learning project he was working on. “I was actually more interested in computer architecture at the time,” Parikh reveals. “But after talking to Dr. Polikar, it sounded like an interesting project.” The experience led Parikh to pursue graduate school in a similar area at Carnegie Mellon University. Here, Parikh cultivated an interest in visual data, which led her down a path to computer vision. Parikh eventually received her PhD in electrical and computer engineering in 2009 working with Professor Tsuhan Chen.
Before joining Facebook, Parikh held positions at various academic institutions, including the Toyota Technological Institute in Chicago, Virginia Tech and Georgia Tech, where she is currently an assistant professor. Through long-term collaborators such as Larry Zitnick, Parikh learned about the environment at Facebook AI Research. When asked about working at Facebook, Parikh remarks the unique level of openness and freedom to work on almost anything. “That is very valuable and very different from a lot of other companies that also do research in AI,” she says. Parikh is based at Facebook HQ in Menlo Park.
In terms of increasing diversity in AI, Parikh points out the cyclical issue of representation and retention in the tech industry. “I think relatability goes a long way […]. If you look at a group of people and you can’t relate to anyone, then it’s just harder,” she says. “You might assume that this isn’t an appropriate fit when that probably isn’t the case.” As women in tech continue to leave their careers at a rate almost twice that of their male counterparts, women entering the field may continue to see a lack of representation. And as Parikh points out, representation helps with retention in the first place. “It’s kind of a chicken and egg problem, to some extent,” she says.
As a result, many diversity initiatives focus on increasing the pipeline of women entering the field. To underrepresented groups considering entering the field of AI, Parikh provides some words of encouragement. “If you’re curious about what’s going on in AI, don’t let that curiosity pass. Follow up on it. For all you know, one thing might lead to another.”
For Adriana Romero Soriano, studying math and science seemed like a familiar and well-trodden path. With a physicist father and biologist mother, going into a STEM field was highly encouraged in the Romero family. “The apple doesn’t fall too far from the tree,” Romero jokes.
Romero recalls first being exposed to AI during her undergrad career, in a collaboration with a PhD student working in computer vision. This collaboration planted a seed of interest in Romero, and from there she eventually went on to pursue a PhD at the University of Barcelona. Her PhD included contributions in the fields of representation learning and model compression, with applications in image classification, image segmentation and remote sensing.
In recounting the path that led her to Facebook AI, Romero highlights her internship experience, where she had the opportunity to work with Yoshua Bengio at Montreal Institute for Learning Algorithms (MILA). “When you’re working with Yoshua—first of all, you can learn a lot. But there is also the fact that by standing on the shoulders of a giant, one can not only see farther but can also be seen, and that’s very important in this community.” Romero says the internship opened a lot of doors, including the opportunity to return to MILA for a postdoc and then transition to Facebook Montreal from there.
Describing her work at Facebook Montreal, Romero mentions the fast-paced, ever-changing environment of Facebook AI Research, where researchers have a great deal of freedom in choosing which projects to pursue. This open approach helped Romero with two papers that will appear at CVPR 2019: Inverse Cooking: Recipe Generation from Food Images and Reducing Uncertainty in Undersampled MRI Reconstruction with Active Acquisition.
Facebook Montreal is also led by Joelle Pineau, whom Romero says she considers a role model. Having role models and fostering a community of like-minded individuals often helps in creating a sense of belonging, especially for underrepresented groups. On the topic of communities, Romero cites organizations such as Women in Computer Vision (WiCV) and Women in Machine Learning (WiML), both of which have several Facebook researchers as active members, sponsors, and participants. “I think [organizations such as WiCV and WiML] are devoting a lot of effort to address these [diversity] challenges, and giving a space to empower women in tech fields by creating a friendly atmosphere for women to engage in technical discussions,” she says. “I personally find this helpful in order to overcome this feeling of not belonging, and I find it very inspiring.”
Foundational models
Latest news
Foundational models