Fairness, Accountability, Transparency and Ethics in Machine Learning

January 9 and 10, 2020

With algorithms based on Machine Learning tending to be ubiquitous in various domains, it is important to ensure that these algorithms are not biased towards any gender, race, ethnicity and other sensitive variables. Recent research trends also point to an increasing interest towards designing such algorithms ensuring fairness.

LinkedIn, Microsoft Research and Indian Institute of Science join hands in organizing a 2-day workshop on Fairness, Accountability, Transparency, and Ethics in Machine Learning. This workshop features tutorials, research talks and hands-on sessions by experts from both industry and academia.

Speakers

Vijay Arya
IBM IRL
Vineeth Balasubramanian
IIT Hyderabad
Siddharth Barman
IISc
Elisa Celis
Yale University
Amit Deshpande
Microsoft Research
Niloy Ganguly
IIT Kharagpur
Sampath Kannan
University of Pennsylvania
Krishnaram Kenthapadi
Amazon
Amit Kumar
IIT Delhi
Jeff Pasternack
LinkedIn
Amit Sharma
Microsoft Research
Nisheeth Vishnoi
Yale University

Panelists

Rushi Bhatt
LinkedIn
Chiranjib Bhattacharyya
IISc
Rahul Matthan
Trilegal
Rohini Srivathsa
Microsoft
Sukriti
LinkedIn

Schedule

The slides for the talks have now been uploaded. Join this LinkedIn events page to access them.

Day 1 (Jan 9th):

Time Event Speaker
08:30 - 14:00
09:15 - 10:30

In this session, we'll discuss what it's like to work on AI problems in industry and how industry differs from academia before diving more deeply into an interactive exploration of how an "Autocomplete" feature would be implemented along with the myriad real-world considerations beyond the core statistical model. "Autocomplete" is a relatively ubiquitous AI-powered writing aid, familiar to users of mobile phones (often, in its simplest form, predicting the completion of the next word as it is being typed) or Gmail (predicting longer phrases to append to the message, accelerating its composition). The statistical models used to implement this have ranged from very basic ngram language models to much more sophisticated recurrent or multi-headed attention (Transformer) encoder-decoder networks; as we'll see, no one approach is strictly best (for example, simpler models tend to be more explainable but more complex models tend to be more expressive), but this is just one of many important, interdependent decisions when building an AI-powered product like Autocomplete. For more details, see this link.

Jeff Pasternack
10:30 - 10:45
10:45 - 13:00

In this session, we'll discuss what it's like to work on AI problems in industry and how industry differs from academia before diving more deeply into an interactive exploration of how an "Autocomplete" feature would be implemented along with the myriad real-world considerations beyond the core statistical model. "Autocomplete" is a relatively ubiquitous AI-powered writing aid, familiar to users of mobile phones (often, in its simplest form, predicting the completion of the next word as it is being typed) or Gmail (predicting longer phrases to append to the message, accelerating its composition). The statistical models used to implement this have ranged from very basic ngram language models to much more sophisticated recurrent or multi-headed attention (Transformer) encoder-decoder networks; as we'll see, no one approach is strictly best (for example, simpler models tend to be more explainable but more complex models tend to be more expressive), but this is just one of many important, interdependent decisions when building an AI-powered product like Autocomplete. see this link.

Jeff Pasternack
13:00 - 14:00
14:00 - 14:10
14:10 - 15:00

We will start with a few real-world scenarios that emphasize the importance of studying fairness, accountability, transparency and ethics in machine learning. We will look at various definitions of individual and group-fairness and what they mean. We will see how different fairness objectives interact with each other and whether fairness comes at some hidden cost. We will do a quick survey of methods to study accountability and transparency of models and individual decisions. In the process, we will see how diverse areas such as ethics, distributive justice, social choice theory, welfare economics etc. have influenced FATE-ML research.

Amit Deshpande
15:00 - 15:45

How do we protect privacy of users when building large-scale AI based systems? How do we develop machine learned models and systems taking fairness, accountability, and transparency into account? With the ongoing explosive growth of AI/ML models and systems, these are some of the ethical, legal, and technical challenges encountered by researchers and practitioners alike. In this talk, we will first motivate the need for adopting a "fairness and privacy by design" approach when developing AI/ML models and systems for different consumer and enterprise applications. We will then focus on the application of fairness-aware machine learning and privacy-preserving data mining techniques in practice, by presenting case studies spanning different LinkedIn applications (such as fairness-aware talent search ranking, privacy-preserving analytics, and LinkedIn Salary privacy & security design), and conclude with the key takeaways and open challenges.

Krishnaram Kenthapadi
15:45 - 16:00
16:00 - 16:45

As machine learning systems are deployed in societally-critical domains like healthcare, governance, and finance, it becomes important to ensure that resultant ML models are both fair across different communities and explanable to key stakeholders. In this talk I will show how fairness and explanation techniques that ignore causality are imperfect, and misleading at worst. Using examples from Bing and other practical scenarios, I will present methods for incorporating causality in reasoning about ML systems. Through these examples, I will argue how fairness and explanation are fundamentally inter-linked.

Amit Sharma
16:45 - 17:30

As machine learning algorithms make further inroads into society, calls are increasing from multiple stakeholders for these algorithms to explain their outputs. Moreover, these stakeholders, whether they be government regulators, affected citizens, domain experts, or developers, present different requirements for explanations. To address these needs, we introduce AI Explainability 360, an open-source software toolkit featuring eight diverse state-of-the-art explainability methods, two evaluation metrics, and an extensible software architecture that organizes these methods according to their use in the AI modeling pipeline. Additionally, we have implemented several enhancements to bring research innovations closer to consumers of explanations, ranging from simplified, more accessible versions of algorithms, to guidance material to help users navigate the space of explanations along with tutorials and an interactive web demo to introduce AI explainability to practitioners. Together, our toolkit can help improve transparency of machine learning models and provides a platform to integrate new explainability techniques as they are developed.

Vijay Arya

Day 2 (Jan 10th):

Time Event Speaker
09:00 - 10:30

Powerful AI systems, which are driven by machine learning (ML) tools, are increasingly controlling various aspects of modern society: from social interactions (e.g., Facebook, Twitter, Google, YouTube), economics (e.g., Uber, Airbnb, Banking), learning (e.g., Wikipedia, MOOCs), to governance (Judgements, Policing, Voting). These systems have a tremendous potential to change our lives for the better, but, via the ability to mimic and nudge human behavior, they also have the potential to be discriminatory, reinforce societal prejudices, and polarize opinions. In this talk, we will give an overview of various notions of fairness in ML and outline our efforts towards controlling bias/discrimination in a principled manner for core machine learning tasks such as classification, data summarization, ranking, personalization, and online advertisement. Our work leads to new algorithms that have the ability to control and alleviate bias from their outputs, comes with mathematical guarantees, and often has low "price of fairness".

Nisheeth Vishnoi, Elisa Celis
10:30 - 10:45
10:45 - 11:30

A number of natural fairness criteria have been proposed for machine learning algorithms. Unfortunately they have been shown to be mutually in conflict. In this talk we will review the criteria and this conflict. We then consider endogenous systems, where people’s behavior and the decision-making algorithm are in equiibrium. We show that there are some notions of fairness that are better aligned with social goals in this setting.

Sampath Kannan
11:30 - 12:15

Major online platforms today (such as Amazon, Netflix, Spotify, LinkedIn, AirBnB) can be thouht of as two-sided markets with producers and customers of goods and services. Traditionally, search and recommendation services in these platforms have focused on maximizing platform usage or customer satisfaction. However, multiple investigations reveal that such customer-centric design of these services may lead to unfair distribution of exposure to the producers and trigger adverse social implications like polarization, unfair representation, popularity bias, unfair distribution of opportunities/business, etc. As more and more people are depending on such platforms for daily news, job/economic opportunities, regular livelihood, etc. it is important to ensure fairness for both producers and customers.

Niloy Ganguly
12:15 - 13:00

K-means is one of the most fundamental clustering problems, and has been widely used for understanding large-data sets. In this talk, I will discuss the so-called constrained k-means problem, where there are additional constraints on the possible clusterings of input data (e.g., data points can come from two different groups, and we may require that they are sufficiently represented in each of the clusters). We give fast algorithms for such problems, which are based on the idea of distance-squared sampling.

Amit Kumar
13:00 - 14:00
14:00 - 14:45

The cake-cutting problem provides a model for addressing fair allocation of a divisible, heterogeneous resource (metaphorically, the cake) among agents with distinct preferences. Focusing on this setup, we will present some of the classic notions of fairness from mathematical economics. I will, in particular, address the guaranteed existence of envy-free (fair) cake divisions and show how this result follows from a combinatorial analog of Brouwer’s fixed-point theorem (Stromquist’80 and Su’99). I will also present some recent results, which complements these existential (and non-constructive) guarantees by way of developing efficient (approximation) algorithms for cake division.

Siddharth Barman
14:45 - 15:30

As neural network (deep learning) models get absorbed into real-world applications each day, there is an impending need to explain the decisions of these neural network models. This talk will begin with an introduction to the need for explaining neural network models, summarize existing efforts in this regard, as well as present a few of our efforts in this direction. In particular, while existing methods for neural network attributions (for explanations) are largely statistical, we propose a new attribution method for neural networks developed using first principles of causality (to the best of our knowledge, the first such). The neural network architecture is viewed as a Structural Causal Model, and a methodology to compute the causal effect of each feature on the output is presented. With reasonable assumptions on the causal structure of the input data, we propose algorithms to efficiently compute the causal effects, as well as scale the approach to data with large dimensionality. We also show how this method can be used for recurrent neural networks. We report experimental results on both simulated and real datasets showcasing the promise and usefulness of the proposed algorithm. This work was presented as a Long Oral at ICML 2019 (http://proceedings.mlr.press/v97/chattopadhyay19a.html).

Vineeth Balasubramanian
15:30 - 16:00
16:00 - 17:00 Rushi Bhatt
Chiranjib Bhattacharyya
Rahul Matthan
Rohini Srivathsa
Sukriti (Moderator)
17:00 - 17:15

Venue

Where

Room: Satish Dhawan Auditorium Maps
Indian Institute of Science,
Bangalore 560012

How to get there

The Institute is known as Tata Institute to the locals. It is better to use the name Tata Institute with the taxi, auto-rickshaw drivers, and bus conductors.

How to attend

Participation in this workshop is only through an invitation.

Organizers

Partha Dutta
LinkedIn
Raman Sankaran
LinkedIn
Namit Chaturvedi
LinkedIn
Devashish Mittal
LinkedIn
Mohit Wadhwa
LinkedIn
Anisha Mascarenhas
LinkedIn
Amit Sharma
Microsoft Research
Amit Deshpande
Microsoft Research
Siddharth Barman
IISc
Chiranjib Bhattacharyya
IISc
Rushi Bhatt
LinkedIn
Satish Sangameswaran
Microsoft Research

Sponsors