Fairness Accountability and Transparency Conference

Nikolaos Sarafianos, February, 26, 2018

I was fortunate to attend the conference on Fairness Accountability and Transparency (FAT*) in NYC, a 2-day multi-disciplinary conference in which ~450 people got together from fields such as social sciences, law, and machine learning. The conference included works on ethics, fairness, justice, interpretability, and transparency and how these intersect with the machine learning models that we build. It was such a refreshing experience since it was the first time I got to talk with people who work in AI policy, with lawyers and social scientists and attend inspiring and thought-provoking keynotes and talks. Below I will try to summarize all the notes I kept and the inspiring things I learned in just two days. You can find the all the proceedings here and some videos here.

TL;DR: FAT* is wonderful, you will have the opportunity to talk with tons of people outside your field, exchange ideas, think of problems that you're working on from a different perspective. Since it's going to take place next year as well I would encourage you to apply, or even better, submit a paper.

Keynotes: The conference started with Latanya Sweeney's keynote in which she made some very compelling arguments on the intersection of privacy and technology. She mentioned that the video cameras of the past did not have a mute button (a technological decision) that enabled people to record things without permission and resulted in them being accused of illegal wiretapping (a law-related decision). One of her main points is that we live in a technocratic era in which the design of new technology leads to laws (and not the other way around) and that technocracy leads to unforeseen consequences with sometimes adverse impact. She gave plenty of examples of ad-targeting algorithms that are biased against specific groups. For example in ads related to arrests, black people have 80% more chances to have the word "arrested" in the ad. In another example, SAT pricing was found to be higher in areas with a high percentage of Asian population. Open questions throughout her talk included how research impacts policy, how and when should regulation respond to tech. When asked about what would be the first thing she would like to see in our algorithms she responded with the transparency and full examination of bias in all parts of the process (data and modeling). The talks will be posted online but in the meantime, there are some similar talks of hers on YouTube (for example the Grace Hopper one) that you can check out.

The second keynote was delivered by Deborah Hellman and discussed what discrimination is, when is it wrong and why. Aiming to distinguish justice and fairness she started her talk with this story: "During campus protests of the 1960s, Sidney Morgenbesser was hit on the head by police. When asked whether he had been treated unfairly or unjustly, he responded that it was "unfair, but not unjust. It was unfair because they hit me over the head, but not unjust because they hit everyone else over the head." She brought examples in which discrimination based on attributes, such as the age in the driving test, is considered fine by everyone although we do not treat everybody equally. Age, after all, is not the perfect proxy for driving she argued and by making such a choice we perform a wrongful discrimination. Her first hypothesis was "Treat like cases alike". She argued that while people are like they are also unlike to one another. For example, she mentioned a NYT article that pointed out that "relative to men with similar criminal histories, women are significantly less likely to commit future violent acts" and thus, if we treated similar cases alike we would end up with harsher decisions for women. She also brought up an excellent point about compounding injustice. For example, imposing higher insurance rates for victims of child abuse which led to an even better question from the audience if this could also include "compounding misfortune". Is treating like cases alike an empty argument she asked. Should it be supplemented by the purpose of the law and assign the punishment that each one deserves? Referencing Peter Westen's Empty Idea of Equality she argued that inaccuracy is a non-comparative notion (given A was charged with X, B can be charged with Y) which is not unfair. Her talk was full of interesting points and raised several questions that have difficult answers.

Talks: A few talks addressed privacy and discrimination in online ads. The paper of Speicher et al. on targeted advertising discrimination discussed how sensitive attributes such as race, age or gender can be used by targeted advertising and that even when such options are not available you can easily find others that are highly correlated with specific groups. For example, anonymized state data can be cross-checked with voting databases in some states which can lead to identifying some people of interest. The work of Datta et al. included a brilliant discussion between a technologist and a lawyer about Google's ad targeting (for a career opportunity customized mainly for men). Let's say that company X puts an ad on Google aiming to target only males and then Google targets mainly males. Is this legal or not they asked? What if the same company wanted the same ad to be directed to all sexes equally but Google's algorithms knew that it should be targeted to men because (for example) most high-earners/CEOs are male. They brought up an argument for responsibility vs capability which translates to lack of control versus lack of incentives. GDPR was a central topic of discussion and it will continue to be in the next few years after it takes effect this spring. I loved the talk on "interventions over predictions" for the topics it raises, the gender stereotyping in Bollywood movies and the 4 steps for revolution they proposed [the first one is watching movies :)]. Finally, the gender shades paper was such a pleasure not only because of the awesome work they did by testing three commercial facial attribute classification algorithms on a balanced dataset in terms of gender and skin color but especially since after its publication, IBM improved their previously biased system to recognize the facial attributes of white males and those of black females equally well. A small comment I have for some papers that include ML algorithms is that they made somewhat larger claims than what they actually had examined since in this conference ablation studies are still absent. Active learning using LIME with a single acquisition function (especially in a batch-mode setup) is a methodology that requires comparisons against more recent AL works (examples here and here) or arguing that by decoupling classifiers (splitting them into 2 groups with each group examining one attribute) results into efficient ML is a big statement without examining how much data you have for each attribute and how imbalanced such attributes could be. Finally, I was told that the workshop on the context and consequences of pre-trial detention was inspiring. A man was invited to tell his story about how his life was affected by algorithmic determination (the video is here). There's also a gofundme page to support him.

Conclusion: I enjoyed FAT* so much more than I had initially expected and I learned tons about the implications of the networks I train on a daily basis and how people from different disciplines think about them. Listening to lawyers discussing the Civil Rights Act and its relation to ad-targeting, social scientists discussing fairness and bias and policymakers thinking what is the right way to approach all these issues made me appreciate so much their work. I'm already looking forward to FAT* 2019.

Acknowledgments: I'd like to thank Solon Barocas and the rest of the organizers for awarding me a travel grant to attend this great conference.

Back to my page