Instructor : Dr. Sreyasee Das Bhattacharjee
Email : sreyasee@buffalo.edu
Office: Davis Hall 349
Office Hours: R, 200pm to 315pm and By Appointment
General Information
Lectures: M W F , 11:00 AM - 11:50 AM
TBD
TA(s):TBD
Course Overview: In this course, we will discuss adversarial learning, analyze explainability as well as the security vulnerability and privacy related issues of different machine learning(ML)/Artificial Intelligence(AI) models, popularly used by the research community. While AI is growingly being employed as an automated decision making tool in several usecase settings like business, education, healthcare, law enforcement, etc., before adopting any such system, it is important for the end users to have a clear understanding of the questions like ‘why the system works?’ than treating it as an omnipotent BlackBox without having any explanation on its trustworthiness. We will review several state-of-the-art research papers to learn about the recent advances in this emerging domain of Trustworthy and Explainable AI, discuss several representative explainable models, learn about different categories of attacks along with a set of certified defenses introduced to evaluate robustness, and finally explore the connections between explainability and trustworthiness in terms of its applications in several domain specific problem settings.
Piazza : We will use Piazza to answer questions and post announcements about the course. Please sign up here.
Course Prerequisites: Experience of Python Programming, Linear Algebra, Calculus, basic understanding of Machine learning/Data mining/Pattern Recognition
Syllabus : You can access the Syllabus here.
Grade Composition : Assignments, 30% Summeries, 15% Mid-term: 20% Final: 25%
Reading Materials:
Several papers will be referred as the reading materials each week, based on the topics we plan to discuss during the week. The details will be shared in Piazza and also be mentioned in the weekly slides.
Other Supporting Materials (will continue to be extended later on):
- Explainable AI: Interpreting, Explaining and Visualizing Deep Learning edited by Wojciech Samek, Gregoire Montavon, Andrea Vedaldi, Lars Kai Hansen, Klaus-Robert Muller, Springer International Publishing : Imprint: Springer; 2019
- Rebooting AI: Building Artificial Intelligence We Can Trust, September 10, 2019, by Gary Marcus (Author), Ernest Davis
- Survey
- Overview of different learning techniques
- Introduction to Adverserial Learning
Course Schedule
(Tentative, more Reading Materials to be added during the course)Week | Lecture | Reading Materials |
---|---|---|
Week 1 | Introduction |
|
Week 2 | Adverserial Learning in Supervised Settings | |
Week 3 | Adverserial Learning in Unsupervised Settings | |
Week 4 | Adverserial Learning in Semisupervised Settings | |
Week 5-6 | Different Types of Attacks: Training Time Attacks and Defenses | |
Week 7-8 | Different Types of Attacks: Test Time Attacks and Defenses | |
Week 9 | MIDTERM | |
Week 10,11 | Attacks by Data Manipulation and Defenses | |
Week 12 | Differential Privacy | |
Week 13 | Understanding and Evaluating Explanability | |
Week 14 | Explaining BlackBox Models | |
Week 15 | Fairness and Ethics |
Academic Integrity:
(Short) Do not cheat! You will be caught and punished. Our department is serious about graduating ethical and upstanding computer scientists. The policy has recently been updated and will be
enforced.
(Long) All academic work must be your own. Plagiarism, defined as copying or receiving materials
from a source or sources and submitting this material as one's own without acknowledging the
particular debts to the source (quotations, paraphrases, basic ideas), or otherwise representing the
work of another as one’s own, is never allowed. Collaboration, usually evidenced by unjustifiable
similarity, is never permitted in individual assignments. Any submitted academic work may be
subject to screening by software programs designed to detect evidence of plagiarism or collaboration. Also, do not post any of the course material outside of the Course piazza page. It will be
interpreted as an attempt to get non-approved help. For more info :
UB CSE Academic Integrity
Working with others: Please do help each other! This material is fun, but can be challenging. Discussing it with peers can deepen your understanding. You can talk about the homework problems and ways of approaching them, however, every person must write up solutions and code separately. We will compare all submissions with each other AND non-approved sources. I you can find something online, so can we.
Special Accommodations: In case of need of special accommodations please go the following link for more information.
Special Accommodations.