Sam Charrington: Today we’re excited to continue the AI for the benefit of society series that we’ve partnered with Microsoft to bring to you. Today we’re joined by Hanna Wallach principal researcher at Microsoft research. Hanna and I really dig into how bias and a lack of interpretability and transparency show up across machine learning. We discuss the role that human biases, even those that are inadvertent, play in tainting data, whether deployment of fair ML algorithms can actually be achieved in practice and much more. Along the way, Hannah points us to a ton of papers and resources to further explore the topic of fairness in ML. You’ll definitely want to check out the show notes page for this episode, which you’ll find at Before diving in I’d like to thank Microsoft for their support of the show and their sponsorship of this series. Microsoft is committed to ensuring the responsible development and use of AI and is empowering people around the world with this intelligent technology to help solve previously intractable societal challenges, spanning sustainability, accessibility and humanitarian action. Learn more about their plan at Enjoy.

Sam Charrington: [00:02:18] All right everyone, I am on the line with Hanna Wallack, Hanna is a principal researcher at Microsoft Research in New York City. Hanna, welcome to this week in Machine Learning and AI.

Hanna Wallach:[00:00:11] Thanks, Sam. It’s really awesome to be here.

Sam Charrington: [00:00:14] It is a pleasure to have you on the show, and I’m really looking forward to this conversation. You are clearly very well known in the machine learning and AI space. Last year, you were the program chair at one of the largest conferences in the field, NeurIPS. In 2019, you’ll be it’s general chair. But for those who don’t know about your background, tell us a little bit about how you got involved and started in ML and AI.

Hanna Wallach:[00:00:48] Sure. Absolutely. So I am a machine learning researcher by training, as you might expect. I’ve been doing machine learning for about 17 years now. So since way before this stuff was even remotely fashionable, or popular, or cool, or whatever it is nowadays. In that time, we’ve really seen machine learning change a lot. It’s sort of gone from this weirdo academic discipline only of interest to nerds like me, to something that’s so mainstream that it’s on billboards, it’s in TV shows, and so on and so forth. It’s been pretty incredible to see that shift over that time. I got into machine learning sort of by accident, I think that’s often what happens. I had taken some undergrad classes on information theory and stuff like that, found that to be really interesting, but thought that I was probably going to go into human computer interaction research.

But through a research assistantship during the summer between my undergrad degree and my Master’s degree, I ended up discovering machine learning, and was completely blown away by it. I realized that this is what I wanted to do. I’ve been focusing on machine learning in various different forms since them. My PHD was specifically on Bayesian Latent Variable methods, typically for analyzing text and documents. So topic models, that kind of thing. But during my PHD, I really began to realize that I’m not particularly interested in analyzing documents for the sake of analyzing documents, I’m interested in analyzing documents because humans write documents to communicate with one another. It’s really that underlying social process that I’m most interested in.

So then during my postdoc, I started to shift direction from primarily looking at text and documents to thinking really about those social processes. So not just what are people saying, but also who’s interacting with whom, and thinking about machine learning methods for analyzing the structure and content of social processes in combination. I then dove into this much more when I got a faculty job, because I was hired as part of UMass Amherst’s Computational Social Science Initiative. So at that point I started focusing really in depth on this idea of using machine learning to study society.

I established collaborations with a number of different social scientists, focusing on a number of different topics. Over the years, I’ve mostly ended up working with political scientists, and often study questions relating to government transparency, and still looking at sort of this whole idea of a social process consists of individuals, or groups of individuals interacting with one another, information that might be used in or arising from these interactions, and then the fact that these things might change over time. I often use one of these or two of these modalities, so structure, content, or dynamics, to learn about one or more of the other ones as well. As I continued to work in this space, I started to think more, not just about how we can use machine learning to study society, but the fact that machine learning is becoming much more prevalent within society.

About four years ago, I started really thinking more about these issues of fairness, accountability, transparency, and ethics. It was a pretty natural fit for me to start moving in this direction. Not only was I already thinking about questions to do with people, but I’ve done a lot of diversity and inclusion work in my non research life. So I’m one of the co-founders of the Women in Machine Learning workshop, I also co-founded two organizations to get more women involved in free and open source software development. So issues related to fairness and stuff like that are really something that I tend to think about a lot in general. So I ended up making sort of this shift a little bit in my research focus. That’s not to say that I don’t still work on things to do with core computational social science, but increasingly my research is focusing on the ways that machine learning impacts society. So fairness, accountability, transparency, and ethics.

Sam Charrington: [00:05:53] We will certainly dive deep into those topics. But before we do, you’ve mentioned a couple of times the term computational social science. That’s not a term that I’ve heard before, I don’t believe. Can you … Is that … I guess I’m curious how established that is as a field, or is it something that is specific to that institution that you were working at?

Hanna Wallach:[00:06:19] Sure. So this is really a discipline that started emerging in maybe sort of 2009, 2008, that kind of time. By 2010, which is when I was hired at UMass, it really was sort of its own little emerging field with a bunch of different computer scientists and social scientists really committed to pushing this forward as a discipline. The basic idea, of course, is you know social scientists study society and social processes, and they’ve been doing this for decades. But often using qualitative methods. But of course, as more of society moves towards digitized interaction methods, and online platforms, and other kinds of things like that, we’re beginning to see much more of this sort of digital data. At the same time, we’ve seen this massive increase, as I’ve said, in the popularity of machine learning and machine learning methods that are really suitable for analyzing data about social processes in society.

So computational social science is really the sort of emerging discipline at the intersection of computer science, the social sciences, and statistics as well. The real goal is to develop and use computational and statistical methods, so machine learning methods, for example, to understand society, social processes, and answer questions that are substantively interesting to social scientists. At this point, there are people at a number of different institutions focusing on computational social science. So yes, of course, UMass, as I’ve mentioned before. But also Northwestern, Northeastern, University of Washington, in fact have been doing this for years, and of course, Microsoft Research is no exception in this regard. Part of the reason why I joined Microsoft Research was that we have a truly exceptional group of researchers in computational social science here. That was really very appealing to me.

Sam Charrington: [00:08:31] Oh, awesome, awesome. So you talked about your transition to focusing on fairness, accountability, transparency, and ethics in machine learning and AI. Can you talk a little bit about what those terms mean to you, and your broader research?

Hanna Wallach:[00:08:54] Yeah, absolutely. So I think the bulk of my own research in that sort of broad umbrella falls within two categories. So the first is fairness, and the second is what I would sort of describe as interpretability of machine learning. So in that fairness bucket, really, much of my research is focused on studying the ways in which machine learning can inadvertently harm or disadvantage groups of people or individual people in various different, usually unintended, ways. I’m interested in understanding not only why this occurs, but what we can do to mitigate it, and what we can do to really develop fairer machine learning systems.

So systems that don’t inadvertently harm individuals or groups of people. In the intelligibility bucket, so there, I’m really interested in how we can make machine learning methods that are interpretable to humans in different roles for particular purposes. There has been a lot of research in this area over the past few years, focusing on oftentimes developing simple machine learning models that can be easily understood my humans simply by exposing their internals, and also on developing methods that can generate explanations for either entire models or the predictions of models. Those models might be potentially very complex. My own work typically focuses really more on the human side of intelligibility, so what is it that might make a system intelligible or interpretable to a human trying to carry out some particular task? I do a lot of human subjects experiments to really try and understand some of those questions with a variety of different folks here at Microsoft Research.

Sam Charrington: [00:11:01] On the topic of fairness and avoiding inadvertent harm, there are a lot of examples that I think many of our audience would be familiar with, the ProPublica work into the use of machine learning systems in the justice process, and others. Are there examples that come to mind for you that are maybe less well known, but that illustrate for you the importance of that type of work?

Hanna Wallach:[00:11:36] Yes. So when I typically think about this space, I tend to think about this in terms of the types of different harms that can occur. I have some work with Aaron Shapiro, Solon Barocas, and Kate Crawford on the different types of harms that can occur. Kate Crawford actually did a fantastic job of talking about this work in her invited talk at the NeurIPS conference in 2017. But to give you some concrete examples, so many of the examples that people are most familiar with are these scenarios as you mentioned where machine learning systems are being used to allocate or withhold resources, opportunities, or information.

So one example would be of the compass recidivism prediction system being used to make decisions about whether people should be released on bail. Another example would be from a story, a news story that happened in November where Amazon revealed that it had abandoned an automating hiring tool because of fears that the tool would reinforce existing gender imbalances in the workplace. So there you’re looking at these existing gender imbalances, and seeing that this tool is perhaps withholding opportunities from women in the tech industry in an undesirable way. There was a lot of coverage about this very sensible decision that Amazon made to abandon that tool. Some other examples would be more related to quality of service issues even when no resources or opportunities are being allocated or withheld.

So a great example there would be the work that Joy Buolamwini and Timnit Gebru did focusing on the ways that commercial gender classification systems might perform less well, so less accurate, for certain groups of people. Another example you might think of, let’s say, speech recognition systems. You can imagine systems that work really well for people with certain types of accents, or for people with voices at certain pitches. But less well for other people, certainly for me. I’m British, and I have a lisp. I know that oftentimes speech recognition systems don’t do a great job of understanding what I’m saying. This is much less of an issue nowadays, but you know, five or so years ago, this was really frustrating for me. Some other examples are things like stereotyping.

So here the most famous example of stereotyping in machine learning is Latanya Sweeney’s work from 2013, where she showed that advertisements that were being shown on web searches for different people’s names would more typically be advertisements that reinforced stereotypes about black criminality when people searched for sort of black sounding names, than when people searched for stereotypically white sounding names. So there the issue is this sort of reinforcement of these negative stereotypes within society by the placement of particular ads for particular different types of searches. So another example of stereotyping in machine learning would be the work done by Joanna Bryson and others at Princeton University on stereotypes in word embeddings.

There has also been some similar work done by my colleague, Adam Kalai, here at Microsoft Research. Both of these groups of researchers showed that if you train word embedding methods, so things like Word2Vec, that try and identify a low dimensional embedding for word types based on the surrounding words that are typically used in conjunction with them in sentences, you end up seeing that these word embeddings reinforce existing gender stereotypes. For example, so the word man ends up being embedded much closer to programmer and similarly woman ends up being embedded much closer to homemaker than vice versa. So that would be another kind of example.

Then we see other kinds of examples of unfairness and harms within machine learning as well. So for example, over and under representation. So Matthew Kay and some others at the University of Washington have this really nice paper where they show that for professions with an equal or higher percentage of men than women, the image search results are much more heavily skewed towards images of men than reality. So that would be another kind of example. What you’ll see from all of these examples that I’ve mentioned is that they affect a really wide range of systems and types of machine learning applications. The types of harms or unfairness that might occur are also pretty wide ranging as well, going from yes, sure, allocational withholding of resources, opportunities of information, but moving beyond that to stereotyping and representation and so on.

Sam Charrington: [00:17:02] So often when thinking about fairness and bias in machine learning and the types of harm that can come about when unfair systems are developed, the kind of all roads lead back to the data itself, and the biases that are inherent in that data. Given that machine learning and AI is so dependent on data, and often much of the data that we have is biased, what can we do about that, and what are the kinds of things that your research is exploring to help us address these issues?

Hanna Wallach:[00:17:41] Absolutely. Yeah, so you’ve hit on a really important point there, which is that in a lot of the sort of public discourse about fairness in machine learning, you have people making comments about algorithms being unfair, or algorithms being biased. Really, I think this misses some of the most fundamental points about why this is such a challenging landscape. So I want to just emphasize a couple of those here in response to your question. So the first thing is that machine learning is all about taking data, finding patterns in that data, and then often training systems to mimic the decisions that are represented within that data. Of course, we know that the society we live in is not fair. It is biased. There are structural disadvantages and discrimination all over the place.

So it’s pretty inevitable that if you take data from a society like that, and then train machine learning systems to find patterns expressed in that data, and to mimic the decisions made within that society, you will necessarily reproduce those structural disadvantages, that bias, that discrimination, and so on. So you’re absolutely right that a lot of this does indeed come from data. But the other point that I want to make is that it’s not just from data and it’s not from algorithms per se. The issue is really, as I see it, and as my colleagues here at Microsoft Research see it, the issue is really about people and people’s decisions at every point in that machine learning life cycle.

So I’ve done some work on this with a number of people here at Microsoft, most recently I put together a tutorial on machine learning and fairness in collaboration with my colleague Jenn Wortman Vaughan. The way we really think about this is that you have to prioritize fairness at every stage of that machine learning lifecycle. You can’t think about it as an afterthought. The reason why is that decisions that we make at every stage can fundamentally impact whether or not a system treats people fairly.

So I think it’s really important when we’re thinking about fairness in machine learning to not just sort of make general statements about algorithms being unfair, or systems being unfair, but really to go back to those particular points and think about how unfairness can kind of creep in at any one of those stages. That might be as early as the task definition stage, so when you’re sitting down to develop some machine learning system, it’s really important to ask the question of who does this take power from, and who does this give power to? The answers to that question often reveal a lot about whether or not that technology should even be built in this first place.

Sometimes the answer to addressing fairness in machine learning is simply, no, we should not be building that technology. But there are all kinds of other decisions and assumptions at other points in that machine learning life cycle as well. So the way we typically like to think about it is that a machine learning model, or method, is effectively an abstraction of the world. In making that abstraction, you necessarily have to make a bunch of assumptions about the world. Some of these assumptions will be more or less justified, some of these assumptions will be better fit for the reality than others. But if you’re not thinking really carefully about what those assumptions are, when you are developing your machine learning system, this is one of the most obvious places that you can inadvertently end up introducing bias or unfairness.

Sam Charrington: [00:21:42] Can you give us some concrete examples there?

Hanna Wallach:[00:21:45] Yeah. Absolutely. One common example of this form would be stuff to do with teacher evaluation. So there have been a couple of high profile lawsuits about this kind of thing. But I think it illustrates the point nicely. So it’s common for teachers to be evaluated based on a number of different factors, but including their student’s test scores. Indeed, many of the methods that have been developed to analyze teacher quality using machine learning systems have really focused predominantly on student’s test scores. But this assumes that student’s test scores are in fact an accurate predictor of teacher quality. This isn’t actually always the case. A good teacher should obviously do more than test prep. So any system that really looks just at test scores when trying to predict teacher quality is going to do a bad job of capturing these other properties. So that would be one example.

Another example involves predictive policing. So a predictive policing system might make predictions about where crimes will be committed based on historic arrest data. But an implicit assumption here is that the number of arrests in an area is an accurate proxy for the amount of crime. It doesn’t take into account the fact that policing practices can be racially biased, or there might be historic over policing in less affluent neighborhoods. I’ll give you another example as well. So many machine learning methods work by defining some objective function, and then learning the parameters of the model so as to optimize that objective function.

So for example, if you define an objective function in the context of, let’s say, a search engine, that prioritizes user clicks, you may end up with search results that don’t necessarily reflect what you want them to. This is because users may click on certain types of search results over other search results, and that might not be reflective of what you want to be showing when you show users a page of search results.

So as a concrete example, many search engines, if you search for the word boy, you see a bunch of pictures of male children. But if you search for the world girl, you see a bunch of pictures of grown up women. These are pretty different to each other. This probably comes from the fact that search engines typically optimize for clicks among other metrics. This really shows how hard it can be to even address these kinds of fairness issues, because in different circumstances the word girl may be referring to a child or a woman, and users search for this term with different intentions. In this particular example, as you can probably imagine, one of these intentions might be more prevalent than the other.

Sam Charrington: [00:24:57] You’ve identified lots of opportunities for pitfalls in the process of fielding systems going all the way back to the way you define your system, and state your intentions, and formulate the problem that you’re going after. Beyond simply being mindful of the potential for bias and unfairness and just saying simply, I realize that that’s not simple, that it’s work to be mindful of this. But beyond that, what does your research offer in terms of how to overcome these kinds of issues?

Hanna Wallach:[00:25:43] Yeah, this is a really good question. It’s a question that I get a lot from people is what can we actually do in practice? There are a number of things that can be done in practice. Not all of them are easy things to do, as you say. So one of the most important things is that issues relating to fairness in machine learning are fundamentally socio-technical. They’re not going to be addressed by computer scientists or developers alone. It’s really important to involve a range of diverse stakeholders in these conversations when we’re developing machine learning systems so that we have a bunch of different perspectives represented. So moving beyond just involving computer scientists and developers on teams, it’s really important that we involve social scientists, lawyers, policy makers, end users, people who are going to be affected or impacted by these systems down the line, and so on and so forth. That’s one really concrete thing you can do.

There is a project that came out of the University of Washington called the Diverse Voices project. It provides a way of getting feedback from stakeholders on tech policy documents. It’s really good, they have a great how-to guide that I definitely recommend checking out. But many of the things that they recommend doing there, you can also think about when you’re trying to get feedback from stakeholders on, let’s say, the definition of a machine learning system. So that task definition stage. Some of these could even potentially be expanded to consider other stages of that machine learning pipeline as well. So there are a number of things that you can do at every single stage of the machine learning pipeline. In fact, this tutorial that I mentioned earlier, that I worked on with my colleague Jenn Wortman Vaughan actually has guidelines for every single step of the pipeline.

But to give you examples, here are some things, for instance, that you can do when you’re selecting a data source. So for example, it’s really important to think critically before even collecting any data. It’s often very tempting to say, oh, there is already some dataset that I can probably repurpose for this. But it’s really important to take that step back and before immediately acting based on availability to actually think about whether that data source is appropriate for the task you want to use it for. There is a number of reasons why it might not be, it could be to do with biases and the data source selection process. There might be societal biases present in the data source itself. It might be that the data source doesn’t match the deployment context, that’s a really important one that people really should be taking into account. Where are you thinking about deploying your machine learning system and does the data you have availability for training and development match that context?

As another example, still related to data, it’s really important to think about biases in the technology used to collect data. So as an example here, there was an app released in the city of Boston back in 2011, I think it was called Street Bump. The way it worked is it used iPhone data and specifically the sort of positional movement of iPhones as people were driving around, to gather data on where there were potholes that should be repaired by the city. But pretty quickly, the city of Boston figured out that this actually wasn’t a great way to get that kind of data, because back in 2011, the people who had iPhones were typically quite affluent and only lived in certain neighborhoods.

So that would be an example about thinking carefully about the technology even used to collect data. It’s also really important to make sure that there is sufficient representation of different subpopulations who might be ultimately using or affected by your machine learning system to make sure that you really do have good representation overall.

Moving onto things like the model, there is a number of different things that you can do there, for instance, as well. So in the case of a model, I mentioned a bit about assumptions being really important. It’s great to really clearly define all of your assumptions about the model, and then to question whether there might be any explicit or implicit biases present in those assumptions. That’s a really important thing to do when you’re thinking about choosing any particular model or model structure. You could even, in some scenarios, include some quantitative notion of parity, for instance, in your model objective function as well. There have been a number of academic papers that take that approach in the literature over the past few years.

Sam Charrington: [00:30:43] Can you give an example of that last point?

Hanna Wallach:[00:30:46] Yeah, sure. So imagine you have some kind of a machine learning classifier that’s going to make decisions of the form, let’s say loan, no loan, hire, no hire, bail, no bail, and so on. The way we normally develop these classifiers is to take a bunch of labeled data, so data points labeled with, let’s say, loan, no loan, and then we train a model, a machine learning model, a classifier, to optimize accuracy on that training data. So you end up setting the parameters of that model such that it does a good job of accurately predicting those labels from the training data. So the objective function that’s typically used is one that considers, usually, only accuracy.

But something else you can do is define some quantitative definition of fairness, some quantitative fairness metric, and then try to simultaneously optimize both of these objectives. So classifier accuracy and whatever your chosen fairness metric is. There is a number of these different quantitative metrics that have been proposed out there that all typically are looking at parity across groups of some sort. So I think it’s really important to remember that even though these are often referred to as fairness metrics, they’re really parity metrics. They neglect many of the really important other aspects of fairness, like justice, and due process, and so on and so forth.

But, it is absolutely possible to take these parity metrics and to incorporate them into the objective function of, say, a classifier, and then to try and prioritize satisfying and optimizing that fairness metric at the same time as optimizing classifier accuracy. There have been a number of papers that focus on this kind of approach, many of them will focus on one particular type of classifier, so like SBMs, or neural networks, or something like that, and one particular fairness metric. There are a bunch of standard fairness metrics that people like to look at. I actually have some work with some colleagues here at Microsoft where we have a slightly more general way of doing this that will work with many different types of classifiers, and many different types of fairness metrics. So there is no reason to start again from scratch if you want to switch to a different classifier or a different fairness metric. We actually have some open source Python code available on GitHub that implements our approach.

Sam Charrington: [00:33:27] So you’ve talked about the idea that kind of people are fundamentally the root of the issue, that these are societal issues, that they’re not going to be solved by technological advancements or processes alone. At the same time, there has been a ton of new research happening in this area by folks in your group and elsewhere. Does that lead to a mismatch between what’s happening in academia and on the technical side with the way this stuff actually gets put into practice?

Hanna Wallach:[00:34:11] That’s an awesome question. The simple answer is yes. This actually relates to one of my most recent research projects, which I’m really, really excited about. So last summer, some of my colleagues and I, specifically Jenn Wortman Vaughan, Miro Dudík, and Hal Daumé, along with our incredible intern, Ken Holstein from CMU, conducted the first systematic investigation of industry practitioner’s challenges and needs for support relating to developing fairer machine learning systems. This work actually came about because we were thinking about ways of developing interfaces for that fair classification work that I mentioned a minute ago. Through a number of conversations with people in different product groups here at Microsoft and people at other companies, we realized that these kinds of classification tasks, while they’re incredibly well studied within the fairness and machine learning literature, are maybe less common than we had thought in practice within industry.

So that got us thinking about whether there might be, actually, a mismatch between the academic literature on fairness and machine learning, and practitioner’s actual needs. What we ended up doing was this super interesting research project that was a pretty different style of research for me and for my colleagues. So I am a machine learning researcher, so is Jen, so is Howell, and so is Miro. Ken, our intern, is an HCI researcher. What we ended up doing was this qualitative HCI work to really understand what it is that practitioners are facing in reality when they try and develop fairer machine learning systems.

To do this, we conducted semi structured interviews with 35 people, spanning 25 different teams, in 10 different companies. These people were in a number of different roles, ranging from social scientist, data labeler, product manager, program manager, to data scientists and researcher. Where possible, we tried to interview multiple people from the same team in order to get a variety of perspectives on that team’s challenges and needs for support. We then took our findings from these interviews, and developed a survey which was then completed by another 267 industry practitioners, again, in a variety of different companies and a variety of different roles.

What we found, at a high level, was that yes, there is a mismatch between the academic literature on fairness in machine learning and industry practitioner’s actual challenges and needs for support on the ground. So firstly, much of the machine learning literature on fairness focuses on classification, and on supervised machine learning methods. In fact, what we found is that industry practitioners are grappling with fairness issues in a much wider range of applications beyond classification or prediction scenarios. In fact, many times the systems they’re dealing with involve these really rich, complex interactions between users and the system.

So for example, chat bots, or adaptive tutoring, or personalized retail, and so on and so forth. So as a result, they often struggle to use existing fairness research from the literature, because the things that they’re facing are much less amenable to these quantitative fairness metrics. Indeed, very few teams have fairness KPIs or automated tests that they can use within their domain. One of the other things that we found is that the machine learning literature typically assumes access to sensitive attributes like race or gender, for the purpose of auditing systems for fairness. But in practice, many teams have no access to these kinds of attributes, and certainly not at the level of individuals. So they express needs for support in detecting biases and unfairness with access only to core screened, partial, or indirect information. This is something that we’ve seen much less focus on in the academic literature.

Sam Charrington: [00:38:41] That last point is an interesting one, and one that I’ve brought up on the podcast previously. In many of the places you might want to use an approach like that, it’s forbidden, from a regulatory perspective, to use the information that you want to use in your classifier to achieve fairness in any part of the decisioning process.

Hanna Wallach:[00:39:04] Exactly. This sets up this really difficult tension between doing the right thing in practice from a machine learning perspective, and what is legally allowed. I’m actually working on a paper at the moment with a lawyer, Zack Conard, actually, a law student, Zack Conard, at Stanford University, on exactly this issue. This challenge between what you want to do from a machine learning perspective, and what you are required to do from a legal perspective, based on humans and how humans behave, and hundreds of years of law in that realm. It’s really challenging, and there is this complicated trade off there that we really need to be thinking about.

Sam Charrington: [00:39:48] It does make me wonder if techniques like or analogous to a differential privacy or something like that could be used to provide a regulatorily acceptable way to access protected attributes, so that they can be incorporated into algorithms like this.

Hanna Wallach:[00:40:07] Yeah, so there was some work on exactly this kind of topic at the FAT ML Workshop colocated with ICML last year. This work was proposing the use of encryption and such like in order to collect and make available such information, but in a way that users would feel as if their privacy was being respected, and so that people who wanted to use that information would be able to use it for purposes such as auditing. I think that’s a really promising approach, although there is obviously a bunch of non trivial challenges involved in thinking about how you might make that a reality. It’s a really complicated landscape. But definitely one that’s worth thinking about.

Sam Charrington: [00:40:54] Was there a third area that you were about to mention?

Hanna Wallach:[00:40:58] Yeah, so one of the main themes that we found in our work studying industry practitioners is a real mismatch between the focus on different points in the machine learning life cycle. So the machine learning literature typically assumes no agency over data collection. This makes sense, right? If you’re a machine learning academic, you typically work with standard data sets that have been collected and made available for years. You don’t typically think about having agency over that data collection process.

But of course, in industry, that’s exactly where practitioners often do have the most control. They are in charge of that data collection or data curation process, and in contrast, they often have much less control over the methods or models themselves, which often are embedded within much bigger systems. So it’s much harder to intervene from a perspective of fairness with the models than it is with the data. We found that really interesting, this sort of difference in emphasis between models versus data in these different groups of people. Of course, many practitioners voiced needs for support in figuring out how to leverage that sort of agency over data collection to create fairer data sets for use in developing their systems.

Sam Charrington: [00:42:20] So you mentioned the FAT ML workshop. I’m wondering as we come to a close, if there are any resources, events, pointers, I’m sure there are tons of things that you’d love to point people at. But what are your top three or four things that you would suggest people take a look at as they’re trying to wrap their heads around this area, and how to either have an impact as a researcher, or how to make good use of it as a practitioner?

Hanna Wallach:[00:42:55] Yeah. Absolutely. So there are a number of different places with resources to learn more about this kind of stuff. So first, I’ve mentioned a couple of times, this tutorial that I put together with Jen Waltman Vahn, that will be available publicly online very soon. It is in fact being broadcast next week, so it should be up by the time this podcast goes live. So I would definitely recommend that people check that out to really get a sense of how we, at Microsoft, are thinking about fairness in machine learning. Then moving beyond that, and thinking specifically on more of the academic literature, the FAT ML workshop maintains a list of resources on the workshop website. That’s again, another really, really great place to look for things to read about this topic. The FAT Star conference is a relatively newly created conference on fairness accountability and transparency, not just in machine learning, but across all of computer science and computational systems. Again, there, I recommend checking out the website to see the publications that were there last year, and also the publications that will be there this year.

There is a number of really interesting papers that I haven’t read yet, but I’m super excited to read, being presented at this year’s conference. That conference also has tutorials on a range of different subjects. So it’s also worth looking at the various different tutorials there. So at last year’s conference, Arvind Narayanan presented this amazing tutorial on quantitative fairness metrics, and why they’re not a one size fits all solution, why there are trade offs between them, why you can’t just sort of take one of these definitions, optimize for it, and call it quits. So I definitely recommend checking that out. Some other places that are worth looking for resources on this, the AI Now Institute, which was co-founded by Kate Crawford, who is also here at Microsoft Research, and Meredith Whitaker, who is also at Google, also has some incredibly awesome resources. They’ve put out a number of white papers and reports over the past couple of years that really get at the crux of why these are complicated socio-technical issues. So I strongly recommend reading pretty much everything that they put out.

I would also recommend checking out some of the material put out by Data and Society, which is also an organization here in New York, led by Danah Boyd, and they too have a number of really interesting things that you can read about these different topics. Then the final thing I want to emphasize is the Partnership on AI, which was formed a couple of years ago by Microsoft and a bunch of other companies working in this space of AI to really foster cross company collaboration and moving forward in this space when thinking about these complicated societal issues that relate to AI and machine learning. So the partnership has been really ramping up over the past couple of years, and they also have some good resources that are worth checking out.

Sam Charrington: [00:46:22] Oh, that’s great. That is a great list that will keep us busy for a while. Hanna, thank you so much for taking the time to chat with us. It was really a great conversation, and I appreciate it.

Hanna Wallach:[00:46:34] No problem. Thank you for having me. This has been really great.

Sam Charrington: [00:46:38] Awesome, thank you.