Welcome to 6.882!

6.882 Spring 2022

Friday, 10:00 AM - 1:00 PM

56-114

Instructor: Dr. Marzyeh Ghassemi

TA: Lelia Marie Hampton

Piazza: piazza.com/mit/spring2022/6882

TA Office Hours: Fridays, 9-10 AM, 56-114

Overview

This course focuses on the human-facing considerations in the pipeline of machine learning (ML) development in human-facing settings like healthcare, employment, and education. Students will learn about the issues involved in ethical machine learning, including framing ethics of ML in healthcare through the lens of social justice. Students will read papers related to ongoing efforts and challenges in ethical ML, ranging from problem selection to post-deployment considerations. Guest lectures will be given by experts in data access, ethics, fairness, privacy and deployments, and the course will focus around a central project that students will use to explore how machine learning can potentially be brought into human-facing deployments ethically.

The course will involve an algorithmic fairness measurement and mitigation exercise and a final project. A major component of this course is the final project. Project presentations will be on the last day of class (May 6).

Students should have taken 2 introductory courses to machine learning; known courses that fulfill these are listed below, but others may be used if similar in content:

Graduate Students: 6.867, 6.806, 6.819
Undergraduate Students: 6.036, 6.401, 6.419, or 6.402

Grading

1. Weekly Reflections: The weekly reflections, corresponding to Week 2 - Week 10, are a total of 10% of your grade.

Done as a Canvas discussion
Must reply to at least one other student
Due before class
Worth 1.25 point per week

2. Problem Sets: Problem sets are worth a total of 85% of your grade, as broken down below.

HW	Subject	Percentage of Grade
1	Algorithmic Fairness Assignment	15%
2	Project Literature Review	15%
3	Project Proposal and Outline	15%
4	Final Project Presentation	20%
5	Final Project Write-Up	20%

Note on Plagiarism: Student code submissions may be submitted by the instructors to a plagiarism detection tool for a review of similarity and detection of possible plagiarism. Submissions will be used solely for the purpose of detecting similarity, and are not retained indefinitely on the server; typically results are deleted after 14 days but may be removed sooner. For more information on the tool used, refer to https://theory.stanford.edu/~aiken/moss/.

3. Scribing: 5% of your grade is assigned for Latexing/scribing the lectures for the week, and submitting them to the class staff for review. Students can sign up in pairs or trios to record the class lectures, and the grade for scribing will be collectively shared by those who share a scribing session.

Scribe LaTeX Template

Scribe Signup

Schedule

Week	Date	Lecture	Materials	HW Due
1/INTRO	2/4	Course Introduction and Overview 10:00 - 10:15 am: Dr. Marzyeh Ghassemi "Course Intro" 10:15 - 11:00 am: Harini Suresh (MIT) "Potential Sources of Harm in ML" 11:00 - 11:10 am: Q/A session --- 11:10 - 12:00 pm: Dr. Marzyeh Ghassemi "Healthcare: Case Study of Hope and Harm" 12:00 - 12:10 pm: Q/A session --- 12:10 - 12:30 pm: Small group interactive case study 12:30 - 1:00 pm: Class discussion	Ethical Machine Learning in Healthcare Understanding Potential Sources of Harm throughout the Machine Learning Life Cycle	HW 1 Out
2/INTRO	2/11	Problem Definition: Disparities in funding and problem selection priorities. 10:00 - 10:50 am: Catherine D'Ignazio (MIT) 10:50 - 11:10 am: Q/A session --- 11:10 - 12:00 pm: Timnit Gebru (DAIR) 12:00 - 12:20 pm: Q/A session --- 12:20 - 1:00 pm: Class discussion Lecture Notes 1 Lecture Notes 2	Data Feminism Introduction (Why Data Science Needs Feminism) Data Feminism Chapter 1 (The Power Chapter) Data Feminism Chapter 4 (What Gets Counted Counts) Datasheet for Datasets Notes on Problem Formulation in Machine Learning	Week 2 Reflection
3/INTRO	2/18	Data Collection: Biased clinical knowledge, power differentials, and social disparities of the healthcare system encode bias. 10:10 - 10:55 am: Emma Pierson (Cornell Tech, MSR) 10:55 - 11:15 am: Q/A session --- BREAK --- 11:20 - 12:20 pm: Class Case Study discussion --- 12:20 - 1:00 pm: Nightingale Data Project Intros Lecture Notes 1 Lecture Notes 2	Quantifying Inequality in Underreported Medical Conditions Race Matters in Health Data Data and its (dis)contents: A survey of dataset development and use in machine learning research Reduced, Reused and Recycled: The Life of a Dataset in Machine Learning Research Mitigating Dataset Harms Requires Stewardship: Lessons from 1000 Papers An algorithmic approach to reducing unexplained pain disparities in underserved populations Artificial intelligence for disparities in knee pain assessment	Week 3 Reflection HW1 Due HW 2 Out
4/INTRO	2/25	Algorithm Development: Default practices, can cause algorithms to work poorly for sub-populations. 10:10 - 11:00 am: Shalmali Joshi (Harvard) 11:00 - 11:15 am: Q/A session --- BREAK --- 11:20 - 11:50 am: Leo Anthony Celi (MIT/Harvard) 11:50 - 12:10 pm: Q/A Session Lecture Notes	Algorithmic Fairness: Choices, Assumptions, and Definitions Ethical limitations of algorithmic fairness solutions in health care machine learning Hidden in Plain Sight — Reconsidering the Use of Race Correction in Clinical Algorithms Fairness and Abstraction in Sociotechnical Systems Fair Regression in Healthcare Spending Sensitivity Analysis of Linear Structural Causal Models Confounding robust policy improvement	Week 4 Reflection
5/INTRO	3/4	Post-Deployment Considerations: Lack of audits of, and adjustment for, shifting populations and use risks poor outcomes. 10:10 - 10:55 am: Aleksander Madry (MIT) 10:55 - 11:15 am: Q/A session --- BREAK --- 11:20 - 12:05 pm: Karen Hao (MIT Tech Review) 12:05 - 12:20 pm: Q/A session --- 12:20 - 1:00 pm: Class Case Study discussion Lecture Notes	Required: How Facebook got addicted to spreading misinformation Skim: A Brief Introduction to Adversarial Examples The Mythos of Model Interpretability Adversarial Examples Are Not Bugs, They Are Features Optional: How Facebook and Google fund global misinformation Can you make AI fairer than a judge? Play our courtroom algorithm game Reproducibility in machine learning for health research: Still a ways to go Do as AI say: susceptibility in deployment of clinical decision-aids Do no harm: a roadmap for responsible machine learning for health care	Week 5 Reflection HW2 Due HW 3 Out
6/FOCUS	3/11	Embodied Data As A Biased Medium: Human data itself is biased due to devices, rules, and societies being biased. 10:10 - 10:55 am: Inioluwa Deborah Raji (UC Berkeley) 10:55 - 11:15 am: Q/A session --- BREAK --- 11:20 - 11:40 pm: Danielle Coleman (U Michigan) 11:40 - 12:20 pm: Q/A session --- 12:20 - 1:00 pm: Class Case Study discussion Lecture Notes	Digital Colonialism: The 21st Century Scramble for Africa through the Extraction and Control of User Data and the Limitations of Data Protection Laws Actionable Auditing: Investigating the Impact of Publicly Naming Biased Performance Results of Commercial AI Products Closing the AI Accountability Gap: Defining an End-to-End Framework for Internal Algorithmic Auditing The Discomfort of Death Counts: Mourning Through the Distorted Lens of Reported COVID-19 Death Data What to Account for When Accounting for Algorithms Reading Race: AI Recognises Patient's Racial Identity In Medical Images Dirty data, bad predictions: How civil rights violations impact police data, predictive policing systems, and justice AI is Learning from Humans. Many Humans.	Week 6 Reflection
7/FOCUS	3/18	Trade-offs in Utility, Privacy and Fairness - There are known trade-offs between utility, algorithmic fairness and privacy. Methods to address biased data via synthetic creation, re-balancing, or algorithmic identification are limited. 10:10 - 11:10 am: Vinith Suriyakumar (MIT) 11:10 - 11:20 pm: Q/A session --- BREAK --- 11:20 - 1:00 pm: Lelia Hampton (MIT) Presentation of SERC Case Study	Chasing Your Long Tails: Differentially Private Prediction in Health Care Settings Can You Fake It Until You Make It? Impacts of Differentially Private Synthetic Data on Downstream Classification Fairness	Week 7 Reflection HW3 Due HW4/5 Out
8	3/25	No Class - Spring Break
9/FOCUS	4/1	Model Robustness and Transfer: Discussion of how model robustness impacts fairness in deployments. 10:00 - 10:50 am: Haoran Zhang (MIT) 10:50 - 11:10 am: Q/A session --- BREAK --- 11:15 - 12:00 pm: Lauren Oakden-Rayner (Australian Institute for ML) 12:00 - 12:20 pm: Q/A session --- 12:20 - 1:00 pm: Class discussion Lecture Notes	Characterizing Generalization under Out-Of-Distribution Shifts in Deep Metric Learning Fairness Violations and Mitigation under Covariate Shift AI for radiographic COVID-19 detection selects shortcuts over signal Feature Robustness in Non-stationary Health Records: Caveats to Deployable Model Performance in Common Clinical Machine Learning Tasks BEDS-Bench: Behavior of EHR-models under Distributional Shift–A Benchmark	Week 9 Reflection
10/FOCUS	4/8	Machine Values: Values that are embedded in machine learning models, and how deployments may alter human systems for better or worse. 10:00 - 10:50 am: Niloufar Salehi (UC Berkeley) 10:50 - 11:10 am: Q/A session --- BREAK --- 11:15 - 12:00 pm: Maxine Mackintosh (Queen Mary University of London) 12:00 - 12:20 pm: Q/A session --- 12:20 - 1:00 pm: Class discussion Lecture Notes	Modeling Assumptions Clash with the Real World: Transparency, Equity, and Community Challenges for Student Assignment Algorithms Do No Harm Measuring Social Biases in Grounded Vision and Language Embeddings Medical Dead-ends and Learning to Identify High-risk States and Treatments Getting Genetic Ancestry Right for Science and Society Defining and Achieving Health Equity in Genomic Medicine Data Representativity for Machine Learning and AI Systems Stability of polygenic scores across discovery genome-wide association studies	Week 10 Reflection
11/CASE	4/15	Case Study: Education 10:00 - 10:50 am: Rene Kizilcec (Cornell) Case Study on Education. 10:50 - 11:10 am: Q/A session --- BREAK --- 11:15 - 12:00 pm: Serena Wang and Lydia T. Liu (Berkeley) 12:00 - 12:15 pm: Q/A session 12:15 - 1:00 pm: Team Quick Hits! 5 Minute Group Presentations	50 Years of Test (Un)fairness Algorithmic fairness in education The many dimensions of algorithmic fairness in educational applications Grade: Machine learning support for graduate admissions Problem formulation and fairness On the limits of algorithmic prediction across the globe How Much Information?: Effects of Transparency on Trust in an Algorithmic Interface Algorithmic Fairness in Education Algorithmic Bias in Education Should College Dropout Prediction Models Include Protected Attributes?
12/CASE	4/22	Case Study: Healthcare 10:00 - 10:50 am: Lauren Wilcox (Google) Case study on Health Deployments 10:50 - 11:10 am: Q/A session --- BREAK --- 11:15 - 12:00 pm: Irene Chen (MIT) Cliffnotes on ethical ML in health 12:00 - 12:15 pm: Q/A session --- 12:15 - 1:00 pm: Team Quick Hits! 5 Minute Group Presentations Lecture Notes	Can AI Help Reduce Disparities in Medical and Mental Health Care? Using search queries to understand health information needs in Africa A Human-Centered Evaluation of a Deep Learning System Deployed in Clinics for the Detection of Diabetic Retinopathy Onboarding Materials as Cross-functional Boundary Objects for Developing AI Assistants Why Is My Classifier Discriminatory?
13/CASE	4/29	Case Study: Employment 10:00 - 10:50 am: Manish Raghavan (Harvard/MIT) 10:50 - 11:10 am: Q/A session --- BREAK --- 11:15 - 12:00 pm: Maria De-Arteaga (UT Austin) 12:00 - 12:15 pm: Q/A Session 12:15 - 1:00 pm Private Project Check-Ins with Course Staff (Link to Book Time)	Mitigating Bias in Algorithmic Hiring: Evaluating Claims and Practices Investigating the impact of gender on rank in resume search engines Bias in Bios Race-Aware Algorithms: Fairness, Nondiscrimination and Affirmative Action Challenges for mitigating bias in algorithmic hiring Auditing employment algorithms for discrimination
14	5/6	Final Project Presentations		HW4 Due
No Class	5/10	No Class		HW5 Due

Project Details

Project Overview

The goal is to have a submission-ready manuscript by the end of the semester, formatted according to the journal the team is targeting to submit to. The project should tackle an ethical issue that could, or has, occurred in the process of using machine learning in a human setting. Projects can be more technical, involving the reproduction or development of models and evaluation, or more socio-technical or policy focused, examining the complex choices, interactions and implications of machine learning use in these spaces.

Teams can be composed of 2-4 students, and there will be one project report/presentation per group. There are many options for the class project but, crucially, students should choose a major undertaking because various components of the project (Project Literature Review, Project Proposal and Outline, Final Project Presentation, Final Project Write-Up) account for a total of 70% of your grade.

Data Sources

Kaggle is a platform for many kinds of data, and competitions from this platform can be modified for relevant investigations.

MIMIC is an open platform for health data. To obtain access to MIMIC, students must obtain CITI certification, and request access on Physionet.

Nightingale is an open platform for health data. Register for the platform as soon as possible, and have a look at existing questions by reading through the documents here. CITI training is required to use the platform. These projects focus mostly on fairness audits of machine learning systems trained on real health data.

Nightengale Project 1: Identifying fairness violations in breast cancer risk using digital pathology images
Every year, 40 million women get a mammogram; some go on to have an invasive biopsy to better examine a concerning area. Since the 1990s, we have found far more ‘cancers’, which has in turn prompted vastly more surgical procedures and chemotherapy. But death rates from metastatic breast cancer have hardly changed. When a pathologist looks at a biopsy slide, they are looking for known signs of cancer: tubules, cells with atypical looking nuclei, evidence of rapid cell division. These features still underlie critical decisions today with respect to how to manage the patient (e.g. give surgery/chemo or wait). Students can predictive models trained on pathology images to predict critical patient outcomes such as mortality and metastasis. They will identify patients at high risk of poor outcomes and compare rates of prediction in different gender/ethnicity groups. https://docs.nightingalescience.org/brca-psj-path.html

Nightingale Project 2: Subtyping cardiac arrest with ECG
A patient is rushed into the ER, unconscious and in cardiac arrest. As the physician begins the resuscitation, they know only that the patient’s heart has stopped—but nothing else. What happened to cause the arrest? What immediate actions need to be taken? One of the only pieces of data available to the emergency physician in this situation is the electrocardiogram (ECG), which measures the electrical activity of the heart. Physicians use this to determine which immediate actions are needed. This rich signal might also contain other clues: about why the heart stopped, what physicians can do in the ER to give the patient the best possible chance of surviving, and the likelihood that a patient who survives will have a normal life, without profound physical or neurological impairments. Students will evaluate the fairness of algorithms using ECGs to predict clinical tasks such as cardiac arrest cause and patient survival post-hospital discharge from hospital. https://docs.nightingalescience.org/arrest-ntuh-ecg.html

Authorship

A note on collaboration: Research is a collaborative activity and we encourage all students to collaborate and learn from each other. In general, when you put your name on something for research, you must: a) have materially contributed to the work, b) be able to defend the research, and c) acknowledge the contribution of others. Keep this in mind when working together and submitting material for evaluation.

A note on authorship: By the end of the course your final project may be sufficiently developed to submit to a peer-reviewed journal. The author order can be a somewhat controversial issue and is left to the project participants to decide. We would strongly encourage you to discuss what the order will be, or what philosophy you will use to decide the order while forming groups. In the case of a dispute during or after the course, the instructors will likely not be able to mediate in any meaningful way. We would also recommend equal authorship (now more common), but the decision is left to each team.

A note on acknowledgement: Papers that result from work done during this course should recognize the contributions of the course in an acknowledgement or in other sections. The suggested language is: "This manuscript was composed by participants in the EECS 6.882 course on Ethical Machine Learning in Human Deployments at the Massachusetts Institute of Technology, Spring 2022.'"

Other MIT Courses on Ethical ML:

Data-Driven Decision Making and Society

Data and Society

STS Computing and Society Concentration

Exploring Fairness in Machine Learning for International Development