GWU_DNSC 6290: Course Outline

Materials for a technical, nuts-and-bolts course about increasing transparency, fairness, security and privacy in machine learning.

• Lecture 1: Self-explainable Machine Learning Models
• Lecture 2: Post-hoc Explanation
• Lecture 3: Discrimination Testing and Remediation
• Lecture 4: Machine Learning Security
• Lecture 5: Machine Learning Model Debugging
• Lecture 6: Responsible Machine Learning Best Practices

Corrections or suggestions? Please file a GitHub issue.

---

Lecture 1: Self-explainable Machine Learning Models

_{^{Source: Simple Explainable Boosting Machine Example}}

Lecture 1 Class Materials

• Syllabus
• Lecture Notes
• Assignment 1
  • Model evaluation notebook
  • Full evaluations results
• Software Examples:
  • Building from Penalized GLM to Monotonic GBM (simple)
  • Building from Penalized GLM to Monotonic GBM
  • Simple Explainable Boosting Machine Example
  • PiML Assignment 1 Example and simple requirements.txt

Lecture 1 Suggested Software

• Python PiML-Toolbox
• Python explainable boosting machine (EBM)/GA2M
• R gam
• h2o penalized GLM (R and Python)
• Monotonic gradient boosting machine (GBM): h2o and xgboost (R and Python)
• R rpart
• Python skope-rules

Lecture 1 Suggested Reading

• Introduction and Background:

  • Designing Inherently Interpretable Machine Learning Models
  • Psychological Foundations of Explainability and Interpretability in Artificial Intelligence
  • Responsible Artificial Intelligence - Sections 2.1-2.5, Chapter 7
  • Stop Explaining Black Box Machine Learning Models for High Stakes Decisions and Use Interpretable Models Instead

• Self-explainaing Machine Learning Techniques:

  • Interpretable Machine Learning - Chapter 5
  • Accurate Intelligible Models with Pairwise Interactions
  • This Looks Like That: Deep Learning for Interpretable Image Recognition

---

Lecture 2: Post-hoc Explanation

_{^{Source: Global and Local Explanations of a Constrained Model}}

Lecture 2 Class Materials

• Lecture Notes
• Software Examples:
  • Global and Local Explanations of a Constrained Model
  • Building from Penalized GLM to Monotonic GBM
  • Monotonic XGBoost models, partial dependence, individual conditional expectation plots, and Shapley explanations
  • Decision tree surrogates, LOCO, and ensembles of explanations
• Assignment 2

Lecture 2 Suggested Software

• Python:
  • allennlp
  • alibi
  • interpret
  • shap
  • PiML-Toolbox
  • tf-explain
• R:
  • DALEX
  • iml

Lecture 2 Suggested Reading

• Introduction and Background:

  • Proposed Guidelines for the Responsible Use of Explainable Machine Learning
  • Responsible Artificial Intelligence - Chapter 4

• Post-hoc Explanation Techniques:

  • Interpretable Machine Learning - Chapter 6 and Chapter 7
  • Towards Better Understanding of Gradient-based Attribution Methods for Deep Neural Networks

---

Lecture 3: Discrimination Testing and Remediation

_{^{Source: Lecture 3 Notes}}

Lecture 3 Class Materials

• Lecture Notes
• Software Example: Testing a Constrained Model for Discrimination and Remediating Discovered Discrimination
• Assignment 3

Lecture 3 Suggested Software

Python:

• aequitas
• AIF360
• Themis

Lecture 3 Suggested Reading

• Introduction and Background:

  • Fairness and Machine Learning - Introduction
  • NIST SP1270: Towards a Standard for Identifying and Managing Bias in Artificial Intelligence
  • Fairness Through Awareness

• Discrimination Testing and Remediation Techniques:

  • Certifying and Removing Disparate Impact
  • Data Preprocessing Techniques for Classification Without Discrimination
  • Decision Theory for Discrimination-aware Classification
  • Fairness Beyond Disparate Treatment & Disparate Impact: Learning Classification Without Disparate Mistreatment
  • Learning Fair Representations
  • Mitigating Unwanted Biases with Adversarial Learning

---

Lecture 4: Machine Learning Security

_{^{Source: Responsible Machine Learning}}

Lecture 4 Class Materials

• Lecture Notes
• Software Example: Attacking a Machine Learning Model
• Assignment 4

Lecture 4 Suggested Software

Python:

• cleverhans
• foolbox
• robustness

Lecture 4 Suggested Reading

• Introduction and Background:

  • A Marauder’s Map of Security and Privacy in Machine Learning
  • BIML Interactive Machine Learning Risk Framework
  • The Security of Machine Learning
  • Proposals for model vulnerability and security

• Machine Learning Attacks:

  • Membership Inference Attacks Against Machine Learning Models
  • Stealing Machine Learning Models via Prediction APIs
  • Model Inversion Attacks that Exploit Confidence Information and Basic Countermeasures
  • Hacking Smart Machines with Smarter Ones: How to Extract Meaningful Data from Machine Learning Classifiers

• Links from Lecture 4:

  • Mitre Adversarial Threat Matrix
  • Robust ML
  • ISIS 'still evading detection on Facebook', report says
  • Researchers bypass airport and payment facial recognition systems using masks
  • Slight Street Sign Modifications Can Completely Fool Machine Learning Algorithms

---

Lecture 5: Machine Learning Model Debugging

_{^{Source: Real-World Strategies for Model Debugging}}

Lecture 5 Class Materials

• Lecture Notes
• Software Example: Debugging a Machine Learning Model
• Assignment 5

Lecture 5 Suggested Software

• DALEX (R)
• SALib (Python)
• themis-ml (Python)
• What-if Tool (Browser)

Lecture 5 Suggested Reading

• Introduction and Background:

  • Debugging Machine Learning Models
  • Why you should care about debugging machine learning models
  • Real-World Strategies for Model Debugging

• Links from Lecture 5:

  • AI Incident Tracker
  • AI Incident Database
  • Debugging Machine Learning Models
  • Underspecification Presents Challenges for Credibility in Modern Machine Learning

---

Lecture 6: Responsible Machine Learning Best Practices

_{^{A Responsible Machine Learning Workflow Diagram. Source: Information, 11(3) (March 2020).}}

Lecture 6 Class Materials

• Lecture Notes
• Assignment 6 (Final Assessment)

Lecture 6 Suggested Software

• Awesome Machine Learning Interpretability

Lecture 6 Suggested Reading

• Introduction and Background:

  • NIST AI Risk Management Framework
  • Interagency Guidance on Model Risk Management (SR 11-7)
  • Eight Principles of Responsible Machine Learning
  • Principles for Accountable Algorithms and a Social Impact Statement for Algorithms
  • Responsible AI Practices

• Links from Lecture 6:

  • Predicting Good Probabilities With Supervised Learning

---

Using Class Software Resources

Mac

Preliminaries:

• Install homebrew: $ /bin/bash -c "$(curl -fsSL https://raw.githubusercontent.com/Homebrew/install/master/install.sh)"

• Install Java:

brew tap adoptopenjdk/openjdk
brew cask install adoptopenjdk8

• Install Mac command line developer tools

• Install Graphviz: $ brew install graphviz

• Install Git: $ brew install git

1. Clone this repository with the examples: $ git clone https://github.com/jphall663/GWU_rml.git

2. Install Anaconda Python 5.1.0 from the Anaconda archives (Anaconda3-5.1.0-MacOSX-x86_64.pkg) and add it to your system path. (This means when you type $ python at the command prompt, you will start a Python 3.6 session. You should probably check for a line like this export PATH="/Users/phall/anaconda/bin:$PATH" in your .bash_profile file. You may also need to close and restart your terminal session.)

3. Install virtualenv: $ pip install virtualenv

4. Change directories into the cloned repository: $ cd GWU_rml

5. Create a Python 3.6 virtual environment: $ virtualenv -p /path/to/anaconda3/bin/python3.6 env_rml (/path/to/anaconda3/bin/python3.6 is not a real path. You must replace it with the path to python3.6 on your system. This can be as simple as $ virtualenv -p python3.6 env_rml.)

6. Activate the virtual environment: $ source env_rml/bin/activate

7. Install the correct packages for the example notebooks: $ pip install -r requirements.txt

8. Start Jupyter: $ jupyter notebook

9. When you are finished, deactivate the environment: $ deactivate

If you get stuck, try this link: Mac Setup: Virtualenv.

Ubuntu

1. Install Git ($ sudo apt-get install git), Graphviz ($ sudo apt-get install graphviz) and Java ($ sudo apt-get install openjdk-8-jdk). They must all be added to your system path.

2. Clone this repository with the examples: $ git clone https://github.com/jphall663/GWU_rml.git

3. Install Anaconda Python 5.1.0 from the Anaconda archives (Anaconda3-5.1.0-Linux-x86_64.sh) and add it to your system path. (This means when you type $ python at the command prompt, you will start a Python 3.6 session.)

4. Install virtualenv: $ pip install virtualenv

5. Change directories into the cloned repository: $ cd GWU_rml

6. Create a Python 3.6 virtual environment: $ virtualenv -p /path/to/anaconda3/bin/python3.6 env_rml

7. Activate the virtual environment: $ source env_rml/bin/activate

8. Install the correct packages for the example notebooks: $ pip install -r requirements.txt

9. Start Jupyter: $ jupyter notebook

10. When you are finished, deactivate the environment: $ deactivate.

---

Some materials Copyright Patrick Hall and the H2O.ai team 2017-2020.