XAI: Learning Fairness with Interpretable Machine

Hey everyone and welcome. My name is Haiyang. I'm a Senior Data Scientist at KPMG Initial Tokyo. Welcome to XAI Learning Fairness with Interpretable Machine Learning. This event is co Hosted by DeepLearning AI, KPMG Initial Tokyo and Machine Learning Tokyo. KPMG Initial Tokyo is a technological hub of KPMG Japan. It develops common digital platform technologies for KPMG's audit, tax and advisory service lines while also supporting the digital transformation of KPMG Japan and our clients. Machine Learning Tokyo is an award winning nonprofit organization dedicated to democratizing machine learning through open education, open source and open science. I'm so excited to be here today facilitating the event representing such international AI community. Today we are honored to have two guest speakers. Our first speaker is Tim Dandy. He's a Director of Board and Chief Solutions Officer at KPMG Initiation Tokyo. Tim will be giving us a talk on KPMG's view on trends in business, adoption of AI and ethical concerns along with our effort at KPMG Indian Tokyo. Our second speaker is Serge Macyz. He's the author of Interpretable Machine Learning with Python and he's a climate and Agronomy Data Scientist at Syngenta. Serge will be giving us a presentation on an overview of many methods employed to detect and mitigate bias and place guardrails to ensure fairness. We will follow with a Q and A session afterwards. Here's a reminder for all our audience today. If you'd like to pose a question for our speakers or upvote others, please find the link to slido in our YouTube description box or in the event reminder email we sent to you. And now, without further ado, let's welcome our first speaker, Tim Danley. Hey Tim, please feel free to take it away.

Thanks Haiyan thanks everybody for giving me an opportunity to talk to you today about some of the things that we're doing at KPMG and some of our thoughts on particularly around ethics and AI. And what I'd like to do is give everybody a bit of a sense of what business leaders are thinking about in terms of AI and ethics in AI. I'm not going to go into a lot of technical details and I'll leave the technical stuff to Serge, who knows much more than I do. But I did want to make sure that we give people that context of the challenge that executives are facing in terms of dealing with AI and these new technologies. Let me share my screen here and start to walk through some of this. First, let me give a quick introduction to my team KPMG Ignition Tokyo so as Hayang mentioned, we're a team that are focused on helping to transform our own business in order to help transform our clients businesses. So we have a team of around 130 technologists, data scientists, software engineers, and we're really focused on helping to drive those latest technologies from Horizon 2 and 3 into our business today. We do this across a number of different things. We obviously have a foundation in technology, but we're also building technology platforms which help our business practitioners to bring their business acumen and their business skills in together with that technology to enhance what we're doing. We're really shifting from professional services based on humans alone to to professionals as a service, using technology as a way to augment and enhance the ability of our technologists to provide that business service. We've been working really closely over the last few years with Machine Learning Tokyo and I think it's a really good partnership in that we've got some interesting knowledge that we can bring to the table in terms of business and what's happening and we're really enjoying working with the network of technologists around the world. So I'm going to start off and say that 2021 is really shaping to be this unprecedented leap of AI into the mainstream. I want to try and give a little bit of perspective across the time frame looking at what I'm seeing in terms of changes in the business context around AI. If I go all the way back to 2017, we surveyed a number of C level executives and what we saw was that a lot of business executives were worried about what data and analytics and AI and these new technologies would do to their reputation. So there was a perception of risk and challenge there, but also that there was a high level of expectation that these new technologies would make a big difference in businesses. At the same time, we surveyed a number of internal auditors and what we discovered was that very few were confident in the governance that was in place around AI. And I think we're only now starting to see the evolution of that and the direct drive and need for regulation around AI. If I fast forward into 2020 with the COVID pandemic around the world, there was a lot of change and evolution in the executive mindset and mandate around AI. The biggest thing was that everybody was beginning to see AI as a way to make their organizations more efficient. And again, increasingly more and more people were thinking that machine learning was the thing that was the biggest impact. At the same time continuing the trend of most companies realizing and understanding that ethics in AI were absolutely necessary, but very few, less than half, had actually implemented anything around a code of ethics for AI. The other key thing is that we were seeing increasingly more business leaders saying that they wanted and expected the government to be involved in regulating AI technology. So many companies are continuing to build on their R and D and reskilling their employees to be able to leverage and use AI. But that still lingering worry of how do you actually make sure that you've got the right governance and control around the AI that you're implementing. The lack of quality in data and the challenges that many organizations, and I'm sure many of you have seen this in the organizations that you work with, really difficult challenges around how to manage and build that quality in data. Moving on to 2021, with the really rapid ramp up of usage of AI in terms of building efficiencies in businesses during the pandemic, we're seeing kind of a doubling down on AI. And I think that when I look at these numbers across a number of different industries, we're absolutely starting to see the shift from this very limited pilot capacity to being functional and even being described as fully functional and at scale in operational context for AI across a number of different industries. However, the perception and the pace of adoption is still quite different across industries. And there is still a perception in many cases that the adoption and usage of AI is proceeding faster than it probably should. So again, bringing back into question those things around the governance and the ethics of AI. Certainly COVID 19 made a huge difference in terms of implementation and accelerated significantly. Most organizations ramp up in usage of AI. However, the confidence levels are growing, but still not at the level where all executives are completely confident in their ability to leverage and completely trust. So that lack of confidence is still there. We see two really critical risks that raise up when we talk to executives about this. The first is cybersecurity breaches, and the second is privacy violations. So that should help give you a bit of context around where organizations are most concerned about the way that they're implementing AI. What I find really interesting about this is that it's not necessarily about ethics. It's fairly traditional understanding of where the risks sit. So I think that we will continue to see and experience that evolution of understanding beyond just the very obvious cybersecurity and privacy aspects. However, there is definitely an understanding that those ethical concerns are are going to continue to grow. And looking at a number of different industries where we would expect to see that is absolutely growing and changing over time. One of the really interesting Things when we talk to industries about this was that smaller and medium businesses are far more likely to say that they have a code of ethics than large companies. And I think this represents, certainly in our experience as kpmg, when we work with very large organizations, that the larger the organization, the harder it is to find ways to actually implement ethical standards. It's relatively easy and there's well established approaches around how to implement compliance and governance standards. But ethics gets into a number of different areas that are harder to judge and then control. The other challenge that I think most executives face, and it's very much represented in this question, was the ability for an organization to keep up with the evolving changes in the AI landscape. And we see an overwhelming majority saying that they are struggling to keep up with and finding it hard to keep up with the landscape of AI. The other thing is this challenge of selecting the right AI technologies. And we often in a business context talk about AI as this broad reaching statement that covers everything. But we all know that AI actually encompasses many, many different types of technologies and approaches to solving problems. And businesses really struggle with wrapping their head around the technology nuances of what those different approaches mean and how they can actually implement those within their business context. So with all of that as context, what I'll say is business leaders are struggling to understand this and business leaders see incredible challenges in terms of ethics in I think the same way that our technical members within organizations are seeing. But there's still not a huge amount of concrete approaches to control and manage and build on those governance constructs. So a couple of key things that we focus on when we talk about ethics and AI and how you implement the first one is integrity. Ensuring that the lineage and appropriateness of how data is used, as well as the genesis of the data itself. Making sure that you have explainability, transparency throughout the process, and ensuring that the decision making process is then articulated in ways that the business can understand naturally. Fairness, ethical, free of bias, free from prejudice. And I know Serge will talk a little bit more about technically some of the ways in which we can do that. And then of course, robustness and resilience in terms of the agility and the compliance that needs to be able to put in place around resistance against bad actors. We have created what we call AI and control, which is a framework around building some of those transparency constructs. Certainly what we've discovered is that when we're able to help the business executives have more confidence through the transparency and frameworks, we're actually enabling the technology teams to be able to move quicker and to be able to move into more advanced technologies. In that context, I won't go through a lot of the details of what KPMG does. I think that's relatively straightforward, but certainly within KPMG niche in Tokyo. One of the key things that we've done is made sure that our focus is not just business. It's making sure that technology across the different pillars of technologies that make sense and are appropriate for a professional services organization are backed up with the technologists required to actually help the executives move through that. So I think that there is a huge amount of potential and there's a huge amount of expectation being built up in terms of executive viewpoint on the usage of AI and these new technologies. And I'm excited to see how we can leverage our ability as technologists to build on and to enhance our various mechanisms to be able to support and help our management to wrap their heads around this new and evolving world of AI. Thanks for giving the opportunity to put some business context on this, and I hope that was helpful. We'll share the slides for everybody so that they can look through some of those details and dig into some of the specifics if they're interested. Thank you.

Thank you, Tim. Thank you very much for your wonderful presentation. I totally agree. Both business and technology side can work together for better explainability and fairness for our audience. If you have any questions for Tim again, you can find the link to slido to post your questions in the YouTube discussion box or in your email reminder email the event reminder email that we sent to you. Okay, so we will be answering those questions after the second presentation. Okay, thank you. Now let's welcome our keynote speaker today, Serge Mercies. Welcome, Serge. The stage is yours.

Hi. Welcome. Greetings from Chicago. And I'll be sharing my slides now. So as soon as I share my slides, make sure you. You put them on. All right. This is my presentation on XAI learning Fairness with interpretable Machine Learning. So first I. I'm going to explain what the session is about. It's broken into five parts. First, we have a brief overview of explainable AI and interpretable machine learning. Then we discuss what fairness is and the origin of bias, followed by the parts about detecting and mitigating bias, which is like the meat of the whole presentation. Lastly, I speculate on what's in store for machine learning and its connection to fairness. Many of the questions asked by attendees when signing up led me to believe that I should start off with what exactly is explainable AI and interpretable machine learning? And how it relates to fairness. There seems to be some confusion here, so I added this section to clarify that so the first thing to know about these terms is that they are the same thing. One stems from academia and another more from industry, but they're used interchangeably. I prefer the term interpretable machine learning, so I'll be using this term more throughout this presentation. The second thing to know is that they have become increasingly more popular over the last five years, as evidenced by these Google trends. Although relatively new in the public domain, these terms have been long present in academic literature. Likewise, during the same period there has been more and more mentions in books for both terms. However, the vast majority of these books have been either non technical books for business audience Books about AI ethics for ethicists, sociologists, psychologists, etc. Or academic compilations for researchers on ML and so on. Practitioner oriented books on the subject have been scarce all these years. My book published only three months ago was the third one. Thankfully, by mid 2020 there'll be six more books on the subject, so definitely there's more buzz about this topic so what exactly is interpretable machine learning? The main concepts are are fairness, accountability and transparency, or fat. You expect the machine learning model to produce fair predictions. They must be robust and traceable back to something or someone, and you should be able to understand how they were made. There are many concerns associated to this concept grounded in ethics, and ultimately what we want to know is can we trust the model? You know, so all these, all these concepts, they just relate to trust. Trust is mission critical. So it's fair to ask, is interpretable machine learning and ethical AI the same thing? Not really, but both concepts overlap. Ethical AI, also known as responsible machine learning, encompasses more than interpretable machine learning because it includes data and algorithmic governance aspects. These aspects include privacy regulations, data management procedures, and even modeling standards. They also uphold fat. On the other hand, interpretable machine learning is how we can implement FAT with machine learning. However, these very same tools, used for ethical reasons, can be used with no ethical reasons in mind, simply for curiosity. They can also be used unethically too. An example is like adversarial robustness. Tools to verify robustness against adversaries can be used for people for nefarious reasons to actually see how they can conduct an attack. I said trust is mission critical, but when you trust something, you're trusting it to fulfill a purpose. I ask, what is the mission of AI? Why do we do all of this? The answer, in my humble opinion, is to enhance or replace human decision making. It's that simple. Those are tall orders, but hopefully not in a bad way, which is why we have to evaluate both options. The both options I'm talking about are human versus model decision making. On one hand, we have a human that carefully takes all these many steps to make a decision right? But this is what Nobel laureate Daniel Kahneman calls System 2 thinking or slow thinking. In reality, most of human decision making is system one thinking or fast thinking. And it's very biased and it can often lead to unfair decisions. On the other hand, we have the model. When we discuss machine learning model decisions, we generally focus on the very last step. Inference is when the model has already been deployed and is being used in real world applications. However, decisions were made before that, not by the model, but by a human. Because of this, there's potential for bias in every prior step. Remember what I said about fast decision making? What part of the problem to focus on, what data to use and the bias that carries how to prepare the data, what model classes to use, what evaluation metrics to use, and how to deploy the model. These are many untraceable decisions that were made before and not communicated or even understood. And they create what is called an accountability gap, which may lead to unfair outcomes. To sum things up, both the model and human decision making are complex and generally opaque. And both have their strengths we can complement and when we use to our advantage. That concludes part one. Now in part two, I will talk about exactly what is fairness, define it, and the origin of bias. What is fairness? There are three notions of fairness. There's treatment, there's process, and there's outcome. Unless robots are interacting directly with us, only the last two pertain to machine learning. First, there's outcomes. In other words, predictions being balanced from a statistical point of view. Then there's the question of having a fair process. Algorithmically, this may mean many things, but I prefer to see it as having no double standards. In other words, different rules apply to similar people. This is what is called a counterfactual notion of fairness. Fairness in practice for machine learning is often a balancing act between procedural and outcome fairness, since fair outcomes seldom come from equal procedures or even balanced procedures. But similar procedures lead to simpler and more intuitive explanations. So explainability and model complexity may be at odds with outcome fairness. And that's something we have to understand about when we were working with fairness. So why focus on fairness? When you hear explainable AI, what comes to mind is explainability and interpretability, which is on the base of this pyramid as transparency. As complicated as this May be this base level in the pyramid is actually the low hanging fruit. And the farther up you go, the harder it becomes. While less on top of mind for machine learning practitioners and more important to the general public. In fact, most of the AI blunders we've seen over the last few years involve issues of fairness and privacy. They don't involve explainability, interpretability issues. So let's begin on the top of the pyramid, where unfairness is manifested as bias in the data and corresponding predictions. We won't be skipping any steps because the top layer depends on the two others. You know, they have to, as a pyramid, it has to hold that layer up. So now what is the origin of bias then? I've already made a hint as far as that concerned. You know that in supervised learning we derive meaning and predictions from a model that's supposed to represent the truth. So in other words, if you see it from the model perspective, the model takes the data and the data represents the truth. Therefore the prediction or whatever insight you take from the prediction coming from the model came from the data which represents the truth. But the most fundamental problem we're aware about is bad data. Typically data entry errors, random if we're lucky. However, there's also biases associated to bad data. We should control for and anomaly detection methods and comparing with external data sets for issues in distribution can help correct these issues. As for the truth, there's a whole host of issues. We have the ugly truths, ugly all the prejudices and systemic biases that exist in the world. And our data is merely depicting. We can make sure that our models mitigate them. Our models don't have to reflect the truth that exists. It can reflect the truth that we want. We also have the fake truths, which are the many ways models can be purposely tricked, and also the truths that change over time. All these issues only aggregate aggravate the whole exercise. So lastly, we have our own biases with our models. Chief among them is observer bias, where we blind ourselves to what we assume or want to be true. Not all these biases lead to unfair outcomes and procedures, but they easily can. Having a data driven approach to machine learning can go a long way in mitigating them. So often when we're talking about explainable AI, people point to complex models. It's easy to plane a complex black box model for bias. But a model, whether black box or not, only amplifies bias that already exists in the data. Therefore, detecting and mitigating bias should be primarily performed to the data and throughout the machine learning workflow to ensure that whatever remains of bias is not amplified. So before you mitigate bias, you have to detect it and measure it. And that's what we're going to discuss in part three. Visualizations provide an easy way to detect bias. Firstly, you have when groups aren't fairly represented in the data set. This is a very common problem that can lead to bias in a model and can be easily visualized by counting how many samples exist for each group. However, it's naive to think that bias is purely representational. Even if you ensure parity between groups, bias may be still stored in the features and their distributions. For instance, the ground truth or predictions might show unfair disparities in outcomes between groups. These disparities can be visualized as you see here in this chart, but they can also through other plots like ROC plots, confusion matrices and contour maps. You may also find troublesome interactions between features that lead to biased outcomes. You can visualize this at heat maps of interaction values or correlations between features. Later on, I'm going to show you an example of how this can be mitigated. Even if it's nice to visualize bias, it's best to quantify it with metrics that measure how well you are minimizing it. For instance, here we have two confusion matrices for two different groups. You can see that one suffers from much more false positives than another. In the context of this example, false positives is a bad thing, so it actually penalizes the group that has the highest false positives. So this could lead you to conclude that the underprivileged group is the one that has in this quadrant a higher number than the other. So this would be underprivileged, and likewise the other one would be the privileged one. From this we can derive certain fairness. First, we have to discuss there are three kinds of fairness metrics. You have the group fairness metrics, you have the individual fairness metrics, and you also have.