Women in Machine Learning: Negar Rostamzadeh
There aren’t enough women in tech. Since the 1980s the number of women completing computer science degrees has plummeted, and in most large tech companies the representation of women in technical roles is below 30%.
This lack of diversity prevents us from building products that work for everybody. It can foster toxic “brogrammer” cultures which harm everybody who works within them, and it deprives teams of the well-documented performance boost that women bring.
Many of the early superstars in computer science were women — from Lord Byron’s polymath daughter Ada Lovelace, the first person to envisage a general purpose computer, to Rear Admiral Grace Hopper, who pioneered the use of natural language in writing computer programs. Similarly, the post-war computing scene was dominated by women. In the words of Grace Hopper: “Programming requires patience and the ability to handle detail. Women are naturals at programming.”
The recent debacle surrounding Google’s James Damore’s memo and subsequent firing illustrates that the source of gender imbalance in tech is still hotly disputed in certain quarters. But many organisations think it’s pretty clear that this gender imbalance stems from current social and cultural norms, and are putting their weight behind the people trying to do something about it.
Negar Rostamzadeh, a fundamental research scientist at Element AI, is one of those toiling away at the grassroots level to help grow the number of successful women in computer science. When she’s not working on deep learning for video classification and generation, she’s organizing this year’s Women in Machine Learning (WiML) workshop at NIPS, Women in Computer Vision (WiCV) at CVPR and Women in Deep Learning (WiDL). We sat down to discuss her role in these organisations and her perspective making our community more diverse and inclusive.
What is Women in Machine Learning?
Negar: In 2005, five women at NIPS [Hanna Wallach, Jennifer Wortman Vaughan, Lisa Wainer, and Angela Yu] realized that they were the only women at the whole conference. They founded WiML to increase the number of women in ML and to enhance the impact of women in the field.
They managed to get Grace Hopper to run their first workshop with 25 attendees, and since then the growth has been massive. At this year’s event in Long Beach there will be 400+ posters!
The impact of WiML has been huge. Firstly, it gives junior academics an opportunity to present their work to senior researchers and get feedback from them. This is really helpful, particularly for newcomers to the field.
Secondly, it makes the field feel like a welcoming place for women. In the words of their mission, WiML aims to “Create opportunities for women (including junior women) in the field to engage in substantive technical and professional conversations in a positive, supportive environment.” This year, for the first time, we had high school students participating in the workshop too. This is exciting because gender imbalance in technical disciplines arises so early that we need to tackle it at source if we want to make sure that there’s a large pool of highly-qualified women to recruit from.
Finally, WiML has provided a template for similar forums in different fields (WiCV, WiDL, and WiNLP) and other inclusive events in tech.
How did you get involved?
I did my undergraduate degree at the University of Tehran, and there were 13 women and 12 men in my CS course. Then I went to a lab in Italy that was also gender balanced. But when I came to North America, I realized that the number of women working in ML was very small.
I attended the 2015 WiML workshop at NIPS, and it was a great way to get to know people in a relaxed atmosphere. These big conferences are very intimidating, particularly if you’re considered a minority, and when you don’t come from a famous lab. Even talking to well-known female researchers is quite tough. Initiating a conversation can be difficult outside of an explicit mentorship setup. The WiML workshops made me feel very comfortable with the other women there, and I found some friends to hang out with for the rest of the conference.
This was my inspiration for co-organizing the third WiCV in 2017. This also led to the creation of WiDL (Women in Deep Learning). One of the women I met at that workshop, Adriana Romero, was actually in my lab (!). We ran WiDL at the Deep Learning Summer School at MILA in 2016 and 2017. So this year, I’ve got my hands full organizing WiML at NIPS, WiCV at CVPR, and WiDL at MILA!
Do you think that North America has more of a WiML problem than, for instance, Iran?
There’s definitely a better ratio in Iran, or rather, in the group of Iranians outside of Iran, because so many great Iranian scientists are now in North America. Eastern European countries are also very balanced; you meet a lot of very impressive female scientists from there.
Doina Precup [McGill University; Google DeepMind] told us at the Canadian Celebration of Women in Computer Science a few months ago: “In Romania nobody told me I wasn’t meant to be a scientist. My grandmother was a mathematician and my mother was a computer scientist. I didn’t know I shouldn’t do computer science until I came to the US.”
This is a strong counterclaim to the idea that there are biological differences between men and women which explain differences in representation. If there are deep-rooted biological differences, why don’t they matter as much in Iran or Romania?
Which female machine learning scientist do you most admire?
There are so many I want to highlight: Hanna Wallach, Kate Crawford, Jennifer Wortman, Anima Anandkumar, Margaret Mitchell.
Timnit Gebru is a formidable researcher and activist. When she sees a problem, she addresses it head-on. She was at NIPS 2016 and she noticed there were very few black participants. This year she organized the first Black in AI workshop at NIPS, and they got around 100 presentations with a ton of senior people.
I also really like Raia Hadsell [Google DeepMind]. She comes from a different academic background (her undergraduate degree is in religion and philosophy) but she was not afraid to switch to a new field where she’s made tremendous contributions.
Here in Montreal, Joelle Pineau [McGill; Facebook] and Doina Precup [McGill; Google DeepMind] are huge inspirations and down-to-earth researchers. They’re at the forefront of Reinforcement Learning and I think their efforts to apply their algorithms to the real world (for instance, predicting seizures) and maximizing the impact of the field are very impressive.
What challenges have you faced in promoting women in machine learning?
Talking about diversity can result in a backlash. Sometimes that’s okay, and we have useful discussions as a result.
The problem emerges when people talk about diversity without actually taking any concrete actions: this generates the backlash and resistance without even producing any positive outcomes. If you talk about diversity a lot but only invite male speakers when organizing a workshop, chances are you’re doing more harm than good. As Rachel Thomas has argued, this “diversity branding” can be very damaging, because it creates the illusion that the problem is solved without helping to alleviate it.
In that sense, the furore surrounding the recent James Damore memo was quite useful. It’s better if people express ideas like this in the open: it reveals the depth of the issue, rather than hiding it. Lots of people say ‘nobody thinks this way any more’; incidents like this show why we need initiatives like WiML, because there is still a sizeable fraction of people who think that women are biologically unsuited to contributing in tech.
How can people get involved?
Firstly, one of the simplest things you can do if you’re organizing a panel or a workshop is to spend some time trying to look outside your circle of friends or collaborators. There are probably female researchers and people from other underrepresented groups who work in the same area but are not very visible. WiML even has a directory to make this easy (if you’re a woman in ML, make sure you’re on this list!). We want to ensure that the research directions that are generated from these meetings are applicable to a wide range of people. Similarly, if you know great female researchers in your field and you like their work, please publicize it!
Secondly, speak up, whether or not you are a minority in this field. If you are invited as a panelist or a speaker to an event with zero or very little diversity, spend some time writing down names of researchers from diverse backgrounds who are experts in the field. Send these names to the organizers and suggest them as future speakers. If you participate unknowingly in an all-male event, point it out afterwards.
Thirdly, if you see a lack of diversity in the place where you work, call it out. If you can, make a female/male salary chart. Companies can’t be forced to do this, but they should — just showing ratios in different departments would be a good start.
Finally, participate in discussions and groups that are dedicated to diversity — regardless of your gender. We have many men in these workshops [WiML, WiDL, WiCV]. We invite everyone to attend!
Further Resources
Negar Rostamzadeh is a Fundamental Research Scientist and Archy de Berker is an Applied Research Scientist, both at Element AI.
Thanks to Marie-Claude Côté and Nasrin Baratalipour for comments.
