Spring 2019 - EE 381V - SPEC TPCS IN MACHINE LEARNING (16725)

---

Course Outline:

• Prerequisites
• Scope
• Instructors
• Schedule and Format

---

Prerequisites

Pre-reqs: At least one graduate course completed in Data Mining/Machine Learning. Online courses do not count.

---

Scope

This is an advanced, seminar-oriented course. We shall study recently published papers relevant to the development of responsible and trustworthy data driven automated decision systems. Solid background in pattern recognition/machine learning is assumed. Key topics include building explainable ML models, black-box explainability, algorithmic fairness, adversarial ML, robust statistical modeling, and privacy aware data mining. Coursework will mainly involve paper presentations, critiques and discussion, a mini coding-based project and a major term project on developing some aspects of a responsible ML system.

---

Instructors

• Instructor: Dr. Joydeep Ghosh
• TA: Diego Garcia-Olano

---

Schedule and Format

Tuesday/Thursday: 12:30 - 2 ECJ 1.312

• Overview 2 classes
• Explainability (P) 7 classes
• Fairness (P) 6 classes
• Assurance (P) 5 classes
• Guest Speaker 4 classes
• Minor project 1 class (mid March)
• Major project 3 classes (late April)

Topics marked by (P) are student-led presentations. Each such class will cover 2 papers, spending 35 minutes per paper as follows: lead group 20 mins, critiquing group, 5 minutes; discussion 10 mins.

Every alternate class marked (P) will include a 5 minute quiz at the beginning of the class.

---

Explainability

Presentation Order	Type	Reading	Year	Venue
1a	interpretable models	Distill-and-Compare: Auditing Black-Box Models Using Transparent Model Distillation	2018	AAAI
1b	interpretable models	Learning qualitatively diverse and interpretable rules for classification	2018	arxiv
2a	black-box explainability	Understanding Black-Box Predictions via Influence Functions	2017	ICML
2b	black-box explainability	Anchors: High-Precision Model Agnostic Explanations	2018	AAAI
3a	black-box explainability	Explanations based on the Missing: Towards Contrastive Explanations with Pertinent Negatives	2018	NIPS
3b	black-box explainability	A Unified Approach to Interpreting Model Predictions	2017	NIPS
4a	black-box explainability	Local Rule-Based Explanations of Black Box Decision Systems	2018	arxiv
4b	neural network oriented	Deep Learning for Case-Based Reasoning through Prototypes: A Neural Network that Explains Its Predictions	2018	AAAI
5a	neural network oriented	Layer-wise relevance propagation for neural networks with local renormalization layers	2016	ICANN
5b	neural network oriented	Streaming Weak Submodularity: Interpreting Neural Networks on the Fly	2017	NIPS
6a	neural network oriented	Rationalizing Neural Predictions	2016	EMNLP
6b	recourse/causal	Actionable Recourse in Linear Classification	2018	arxiv
7a	philosophy	The mythos of model interpretability	2016	ICML
7b	philosophy	Towards a rigorous science of interpretable machine learning	2017	arxiv

Fairness

Presentation Order	Type	Reading	Year	Venue
8a	toolkit	AI FAIRNESS 360	2018	Arxiv
8b	bias detection	Fast Threshold Tests for Detecting Discrimination	2018	AISTATS
9a	formalism	On formalizing fairness in prediction with machine learning	2018	ICML
9b	pre-processing	Optimized Pre-Processing for Discrimination Prevention	2017	NIPS
10a	in-processing	Mitigating Unwanted Biases with Adversarial Learning	2018	AAAI
10b	in-processing	Classification with Fairness Constraints: A Meta-Algorithm with Provable Guarantees	2019	ACM
10b	history	50 Years of Test (Un)fairness: Lessons for Machine Learning	2019	FATML
11a	post-processing	Equality of Opportunity in Supervised Learning	2016	NIPS
11b	post-processing	On Fairness and Calibration	2017	NIPS
12a	causal	Causal Reasoning for Algorithmic Fairness	2018	arxiv
12b	temporal	Delayed Impact of Fair Machine Learning	2018	ICML
13a	ranking	Fairness of Exposure in Rankings	2018	KDD
13b	ranking	Towards a Fair Marketplace: Counterfactual Evaluation of the trade-off between Relevance, Fairness & Satisfaction in Recommendation Systems	2018	ACM
14a		Controlling Polarization in Personalization: An Algorithmic Framework	2019	FATML

Assurance

Presentation Order	Type	Reading	Year	Venue
15a	data integrity	Datasheets for Datasets	2018	PMLR
15b	data integrity	Mitigating poisoning attacks on machine learning models: A data provenance based approach	2017	ACM
16a	overview	Making Machine Learning Robust Against Adversarial Inputs	2018	ACM
16b	causality	Reliable Decision Support Using Counterfactual Models	2017	NIPS
17a	tool	ART sec 1-5 Nicolae, Maria-Irina, et al. 2018. “Adversarial Robustness Toolbox v0.3.0.”	2018
17b	tool	ART sec 6-10	2018
18a	DL specific	Towards Deep Learning Models Resistant to Adversarial Attacks	2018	ICLR
18b	DL specific	Decomposition of uncertainty in Bayesian deep learning for efficient and risk-sensitive learning	2018	ICML
19a	outliers	Generative Adversarial Active Learning for Unsupervised Outlier Detection	2018	arxiv
19b	outliers	Theoretical Foundations and Algorithms for Outlier Ensembles	2015	SIGKDD
20a	bias mitigation	Uncovering and Mitigating Algorithmic Bias through Learned Latent Structure	2019	AAAI
20b	bias mitigation	Why is my classifier Discriminatory?	2018	NIPS
21a	security/privacy	Model Reconstruction from Model Explanations	2019	FATML
21b	software process/evaluation	Model Cards for Model Reporting	2019	FATML