As machine learning and deep learning models are impacting on real world environments (e.g., ChatGPT), concerns on the trustworthiness of machine-learned models are rising. In this course, we explore whether popular machine-learned models are trustworthy and then study various learning methods to enchant the models to be trustworthy. To this end, we will learn basic knowledge on machine learning theory, uncertainty learning via conformal prediction, adversarial examples/learning, machine unlearning, differentially private learning, fairness in learning, and miscellaneous topics on trustworthy generative AI.

- 선수과목: 확률및통계
- 필수과목: 기계학습

Discussion: 50%
Final presentation: 50%

- Learning theory: Foundation of Machine Learning (https://cs.nyu.edu/~mohri/mlbook/)
- Conformal prediction: A Tutorial on Conformal Prediction (https://arxiv.org/pdf/0706.3188.pdf)
- Differential Privacy: The Algorithmic Foundations of Differential Privacy (https://www.cis.upenn.edu/~aaroth/Papers/privacybook.pdf)
- Fairness: FAIRNESS AND MACHINE LEARNING (https://fairmlbook.org/pdf/fairmlbook.pdf)

Tentative Schedule
[Week 1-3] Preliminaries on machine learning theory
- Introduction
- Concentration inequalities
- Probably approximately correct (PAC) learning under finite hypotheses and realizable case
- Learning under infinite hypotheses and unrealizable case
- Statistical query (SQ)
- Online learning
[Week 4-5] Uncertainty Learning: Conformal Prediction
- Introduction
- Conformal prediction
- PAC conformal prediction and its applications
- Adaptive conformal prediction and its applications
[Week 6-7] Adversarial Example and Learning
- Introduction
- Adversarial example generation
- Heuristic adversarial learning
- Certified adversarial learning
[Week 8] Machine unlearning
- Introduction
- SQ review
- Unlearning with SQ
[Week 9-10] Differentially private learning
- Introduction
- differential privacy (DP)
- DP in SQ
- Applications of DP in large language models (LLMs)
[Week 11-12] Fairness in learning
- Introduction
- Fairness definitions
- Fairness in practice
[Week 13-14] Miscellaneous topics on trustworthy generative AI
- Vulnerabilities in generated code
- Prompt tuning
- Copyright issue
[Week 15-16] Final presentation