3 Highlights from FAT*ML 2020
Did you miss this year’s FAT*ML conference? No worries, here comes my top 3 takeaways to keep you updated on some of the main hot topics currently being discussed in the FAT*ML community.
FAT*ML is an interdisciplinary conference that brings together researchers and practitioners interested in fairness, accountability, and transparency in socio-technical systems. Although the FAT*ML community started with a strong focus on algorithms and Computer Science, this year’s conference showed a noticeable shift towards real interdisciplinary research, extending its focus within law, social sciences, and humanities. This is reflected by the number of accepted papers from authors belonging to different disciplines. But why should we even care about fairness, accountability, and transparency in machine learning? In case you’re not already familiar with the ideas and thoughts behind the FAT*ML acronym, please take a look at Marco Angel Bertani-Økland’s blog post introducing the topic before last year’s conference.
1. Is our anti-discrimination focus right?
The first theme from this year’s conference I found particularly interesting and important is the shift towards a broader perspective on fairness. Previously, research within the FAT*ML community has been primarily focusing on fairness from a perspective of non-discrimination and measurable equality in results between groups or individuals. Input from other disciplines has this year expanded this rather narrow focus of fairness and opened for new questions addressing issues such as power structures, inclusive processes and the role of researchers and scientists in problem formulation and solution approaches.
Within the community, the assumption of a general interest in fairness and fair results seems often to be generally accepted. It was however stressed multiple times that this assumption might often be wrong and that what is fair is often defined by the developers or owners of a system rather than the users. As such, the result is that the providers decide what people value. Different talks addressed these unbalanced power issues and suggested many new and interesting approaches and methodologies on how to achieve greater fairness and equality in machine learning-based systems. Interesting talks include among others: Michael Katell’s talk on how to achieve equality through the process over the product through early and often communication with those affected by machine learning, Chelsea Barabas’ talk on the advantages and methodologies in studying the powerful rather than the powerless and Bogdan Kulynych’s contribution on how to design and implement a new class of defenses to enable those affected by optimization systems to influence, alter, and counter systems from the outside.
The existing very narrow view on fairness from within the community was also contested by a very inspirational keynote talk by Yochai Benkler. In his talk, he focused on his conclusions on the two sides of fairness: discrimination and exploitation. He reached out to the community and asked for a shift in focus from anti-discrimination (uneven treatment in outcomes) to anti-exploitation (uneven distribution of power and resources), to address a broader range of potential harms considering productivity and power distribution. As such he pointed at larger structural issues of the relation between social status or class, power, and exploitation of humans and emphasized the potential in the computer science disciplines to help address exploitation issues. If we were able to better measure exploitation in an objective manner inspired by how we evaluate discrimination, this community could have a stronger impact on the world.
2. Are explanations useful?
Another interesting input from this year’s conference was the talks and panel discussion addressing the use and value of explanations for end users. One main reason behind research in machine learning explanations is to achieve greater transparency and empower the user to understand what to do to change his or her outcome from the algorithm in the future and to be able to criticize and question the system. However, in research presented by Umang Bhatt, a key finding is that most often when explanations are used in practice, they never reach the end-user but are rather used for debugging by the data engineer. Furthermore, this research points out some technical limitations related to real-time deployment, as well as issues related to how to interpret explanations and the uncertainty behind relations and causality in the suggested explanations. One talk by Solon Barocas also pointed out that explanations rely on a number of easily overlooked assumptions, including that the recommended change in feature values clearly maps to real-world actions. Another talk by Leif Hancox-Li also addressed the issue that explanations are often non-robust and argues how this is problematic if we wish to explain real patterns in the world and how this can constitute a moral hazard where the “best” explanations are chosen based on which explanations look most fair.
Acknowledging some of these limitations in explanations led to a discussion regarding the usefulness of explanations for end-users, and if explanations are based on-premises and technical and statistical assumptions that users do not understand. This also raises the question of whether we should use more time in building interpretable models instead of using black-box models (see section 7 on the fair-washing risk of explanations). Related to this question Ravit Dotan and Smitha Milli presented some interesting work pointing out that the choice of the machine learning model is not value-neutral. They pointed to the fact that the use of neural networks and deep-learning models are correlated with the increase in computing power, but that these models have some trade-offs including environmental impacts, centralization of power and lack of interpretability.
3. How to do ethics in practice?
Finally, this year’s conference also focused a great deal of attention on the issues of regulation and internal auditing. Quite some papers considered fairness, transparency, and explanations in the light of GDPR, including a keynote speech by Nani Jansen Reventlow.
But with still very few regulations and official audit systems or institutions in place, internal auditing and self-regulation is something many practitioners and corporate companies are trying to tackle at the moment. To guide this process Inioluwa Deborah Raji and Andrew Smart presented their work on defining an End-to-End Framework for Internal Algorithmic auditing (before deployment). Related to this, and as also mentioned above, Bogdan Kulynych also introduced an interesting framework for a broader external audit of machine learning systems after deployment.
Also, an interesting craft session on how to bridge the gap from AI Ethics Research to practice focused on how in practices some of the central tech companies such as Facebook, LinkedIn, and Pymetrics (automated hiring) are working with ethics and fairness internally. In a discussion group afterward the need for proper regulation and standards was discussed along with the importance of ensuring that both internal practices and external audits do not just result in ethics washing, but actual ethical considerations.
Of course, many other interesting talks and questions occurred during the conference, but these were some of the key take-aways that I found particularly interesting. To get an overview of all workshops and talks, you can find the whole program here and all accepted papers here.
