Term: Aug - Dec’ 24
Time: Tuesdays & Thursdays (11:30 - 13:00)
Venue: CDS 419
Credits: 3:0
Course Details
Outline: We interact with AI technology on a daily basis—such systems answer the questions we ask (using Google, or other search engines), curate the content we read, unlock our phones, allow entry to airports, etc. Further, with the recent advances in large language and vision models, the impact of such technology on our lives is only expected to grow. This course introduces students to ethical implications associated with design, development and deployment of AI technology spanning NLP, Vision and Speech applications.
This is a seminar-style course, wherein each class would be a discussion based on the readings assigned for that particular day. Each class would begin with a short quiz, which would be straightforward if you have read through the required section of the reading material. The in-class discussion among students would be facilitated by the instructor who would bring in discussion points based on the reading material.
Prerequisites: The class is intended for graduate students and senior undergraduates. Students should have finished at least a basic machine learning course (from IISc), and any one IISc course related to the discussed applications (computer vision, speech or NLP).
Content: Specifically, this seminar course would facilitate discussions among students structured around pre-selected readings on topics related to ethics in AI. Specifically, we plan to read about and discuss following modules:
- M1. Overview of ethical theories
- M2. Data collection and curation
- M3. Biases and algorithmic fairness; debiasing and mitigating harms
- M4. Privacy
- M5. Content Moderation
- M6. Misinformation, disinformation and hate-speech
- M7. Algorithm audits and transparency
- M8. Environmental impact of model training and inference
- M9. Future of work; economic impact of AI
Intentionally, many of the modules are connected to each other. We hope to spend at least 2-3 classes discussing each module.
Schedule:
A tentative schedule for the course is below.
| Date |
Topic |
Reading Material |
| Aug 6 |
Course Overview |
– |
| Aug 8 |
(M1) Major Ethical Theories |
Required: (a) An overview of ethical theories |
| Aug 20 |
(M1) Values in NLP/ML research |
Required: (a)The Social Impact of NLP; and (b) The Values Encoded in ML Research (pages 1-15) |
| Aug 22 |
(M2) Data Cascades |
Required: “Everyone wants to do the model work, not the data work" |
| Aug 29 |
(M2) Data documentation |
Required: Datasheets for datasets (pages 1-10). Recommended: Data Statements for NLP |
| Sep 3 |
(M2) Data collection |
Required: IndicVoices: Towards building an Inclusive Multilingual Speech Dataset for Indian Languages. Recommended: IndicVoices Blog, Ethical Data Pledge, OpenAI Used Kenyan Workers on Less Than $2 Per Hour to Make ChatGPT Less Toxic |
| Sep 5 |
(M3) Algorithmic Bias |
Required: Moving beyond ‘‘algorithmic bias is a data problem’’. Recommended: Facial Recognition Is Accurate, if You’re a White Guy, The Woman Worked as a Babysitter: On Biases in Language Generation |
| Sep 10 |
(M3) Algorithmic Bias |
Required: Language (Technology) is Power: A Critical Survey of “Bias” in NLP (pages 1-9). Recommended: Gender Bias in Coreference Resolution |
| Sep 12 |
(M3) Debiasing and mitigation |
Required: Mitigating Gender Bias in NLP. Recommended: Getting Gender Right in Neural Machine Translation |
| Sep 17 |
(M4) Privacy |
Required: What Does it Mean for a Language Model to Preserve Privacy? Recommended: Differential Privacy: A Primer |
| Sep 19 |
Project Discussions |
– |
| Oct 1 |
(M4) Privacy |
Required: Extracting Training Data from Large Language Models, What does GPT-3 “know” about me? |
| Oct 3 |
(M5) Content Moderation |
Required: When Curation Becomes Creation: Algorithms, Microcontent, and the Vanishing Distinction between Platforms and Creators. |
| Oct 8 |
(M5) Content Moderation |
Required: Content moderation, AI, and the question of scale. Recommended: Do Not Recommend? Reduction as a Form of Content Moderation |
| Oct 15 |
(M5) Content Moderation |
Required: Decolonizing Content Moderation. Recommended: AI Content Moderation, Racism and (de)Coloniality |
| Oct 17 |
Peer feedback on project proposal |
– |
| Oct 22 |
(M6) Misinformation, disinformation |
Required: Images and misinformation in political groups: Evidence from WhatsApp in India and Tiplines to Combat Misinformation on WhatsApp. Recommended: Can WhatsApp Benefit from Debunked Fact-Checked Stories to Reduce Misinformation? |
| Oct 24 |
(M6) Misinformation, disinformation |
Required: A survey on automated fact-checking. Recommended: How to search for fact-checked information |
| Oct 29 |
(M7) Algorithm Audits |
Required: Auditing Algorithms Understanding Algorithmic Systems from the Outside: Chapter 2, 3 (also recommend reading 1 and 4). |
| Oct 31 |
No class (Diwali break) |
– |
| Nov 5 |
(M7) Algorithm Audits |
Required: Unequal Representation and Gender Stereotypes in Image Search Results for Occupations. Recommended: An Image of Society: Gender and Racial Representation and Impact in Image Search Results for Occupations, Text-to-Image Generation Amplifies Demographic Stereotypes at Large Scale |
| Nov 8 |
(M8) Environmental Impact |
Required: Environmental Section of Foundations Models Paper (pages 140 - 145) and Power Hungry Processing: Watts Driving the Cost of AI Deployment?. Recommended: Estimating the Carbon Footprint of BLOOM,a 176B Parameter Language Model, Systematic Reporting of the Energy and Carbon Footprints of Machine Learning, Green AI: 1, 2, |
| Nov 14 |
(M9) Future of work |
Required: Thousands of AI Authors on the Future of AI |
Course Evaluation
The evaluation comprises 3 components:
- Class projects [35%]: we will shortly share more details about the course projects
- Readings & discussion [35%]: A small group would be responsible for scribing one module (constituting 20%), the in-class quizzes account for 10% of the grade (we would drop the lowest two scores). A small 5% of the grade is reserved for in-class participation.
- Final exam [30%].
Scribes
Each course attendee would be responsible for scribing one module of the course.
A high-quality scribe (corresponding to A, A- score) would comprehensively discuss the class interactions
in relation with the required and recommended reading material.
Less comprehensive and less thoughtful scribes would warrant lower scores
(for instance, summarizing just the class discussions or the reading material but not both).
Class Projects
The course project constitutes 35% of the overall score, where students (with team sizes of no more than two) get a chance to apply the acquired knowledge. So long as the broad topic of the project is relevant to the course,
projects could take the diverse forms, including case studies, qualitative research projects, model and/or data audits, algorithmic inquiries and solutions, etc. The project includes two milestones: (1) initial proposal, which clearly states the problem, discusses relevant literature and outlines a rough action plan; and (2) a final report. Towards the end of the course, students would get a chance to showcase their research through presentations.
Each team would get two late days for projects, no extensions will be offered (please don’t even ask). After your late days expire, you can still submit your project but your obtained score would be divided by 2 if submitting after 1 day, and will be divided by 4 if submitting after 2 days. No submissions would be entertained after that.
Important Dates:
- Sept 30, 16:59 Project proposals due
- Nov 21, 16:59 Final reports due
Discussions & (Anonymous) Feedback
We will use Teams for all discussions on course-related matters. Registered students will receive the joining link/passkey.
If you have any feedback, you can share it (anonymously or otherwise) through this link: http://tinyurl.com/feedback-for-danish
Teaching Staff
- Kinshuk Vasisht (Teaching Assistant)
- Danish Pruthi (Instructor)
Acknowledgements
We are grateful to Yulia Tsvetkov for readily sharing the content for her ethics classes at CMU and UW, and to Navreet Kaur for helping with the initial outline of this course.
-