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: Interpretable 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: Interpretable Machine Learning Models

_{^{Source: Building from Penalized GLM to Monotonic GBM}}

Lecture 1 Class Materials

• Syllabus
• Lecture Notes
• Lecture Video
• Software Example: Building from Penalized GLM to Monotonic GBM

Lecture 1 Suggested Software

• 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:

  • Stop Explaining Black Box Machine Learning Models for High Stakes Decisions and Use Interpretable Models Instead
  • Responsible Artificial Intelligence - Sections 2.1-2.5, Chapter 7

• Interpretable Machine Learning Techniques:

  • Interpretable Machine Learning - Chapter 4
  • Accurate Intelligible Models with Pairwise Interactions
  • This Looks Like That: Deep Learning for Interpretable Image Recognition

• Links from Lecture 1:

  • Tay (bot)
  • New York Regulator Probes UnitedHealth Algorithm for Racial Bias
  • When a Computer Program Keeps You in Jail
  • When an Algorithm Helps Send You to Prison

---

Lecture 2: Post-hoc Explanation

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

Lecture 2 Class Materials

• Lecture Notes
• Lecture Video
• Software Example: Global and Local Explanations of a Constrained Model

Lecture 2 Suggested Software

• Python:
  • allennlp
  • alibi
  • interpret
  • shap
  • 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 5 and Chapter 6
  • Towards Better Understanding of Gradient-based Attribution Methods for Deep Neural Networks

• Links from Lecture 2:

  • On the Art and Science of Explainable Machine Learning
  • Access Denied: Faulty Automated Background Checks Freeze Out Renters
  • ML Attack Cheatsheet
  • Debugging Machine Learning Via Model Assertions
  • Machine Bias
  • Gender Shades
  • Explainable Neural Networks based on Additive Index Models

---

Lecture 3: Discrimination Testing and Remediation

_{^{Source: Lecture 3 Notes}}

Lecture 3 Class Materials

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

Lecture 3 Suggested Software

Python:

• aequitas
• AIF360
• Themis

Lecture 3 Suggested Reading

• Introduction and Background:

  • Fairness and Machine Learning - Introduction
  • 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:

_{^{Source: Secure Machine Learning Ideas}}

Lecture 4 Class Materials

• Lecture Notes
• Lecture Video
• Software Example: Attacking a Machine Learning Model

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:

  • A Plea for Simplicity
  • Privacy Risks of Explaining Machine Learning Models

---

Lecture 5: Machine Learning Model Debugging

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

Lecture 5 Class Materials

• Lecture Notes
• Lecture Video
• Software Example: Debugging a Machine Learning Model

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:

  • Testing and Debugging in Machine Learning (Google)
  • AI Incidents (not already linked above):
    • Self-Driving Uber Car Kills Pedestrian in Arizona, Where Robots Roam
    • The Woz tweets on Apple and Goldman Sachs
    • Suckers List: How Allstate’s Secret Auto Insurance Algorithm Squeezes Big Spenders
    • A.C.L.U. Accuses Clearview AI of Privacy ‘Nightmare Scenario’
    • Government’s Use of Algorithm Serves Up False Fraud Charges
    • Microsoft's robot editor confuses mixed-race Little Mix singers
    • Welfare surveillance system violates human rights, Dutch court rules
  • Intelligible Models for HealthCare: Predicting Pneumonia Risk and Hospital 30-day Readmission

---

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
• Lecture Video

Lecture 6 Suggested Software

• Awesome Machine Learning Interpretability

Lecture 6 Suggested Reading

• Introduction and Background:

  • Eight Principles of Responsible Machine Learning
  • Principles for Accountable Algorithms and a Social Impact Statement for Algorithms
  • Responsible AI Practices

• Links from Lecture 6:

  • Example Model Card
  • Network Graph Example
  • Autoencoder Visualizations:
    • Reducing the Dimensionality of Data with Neural Networks
    • DNSC 6279 Autoencoder Example

---

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 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.