GWU_DNSC 6330: Course Outline

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

• Lecture 1: Explainable Machine Learning Models
• Lecture 2: Post-hoc Explanation
• Lecture 3: Bias Testing and Remediation
• Lecture 4: Machine Learning Security
• Lecture 5: Machine Learning Model Debugging
• Lecture 6: Responsible Machine Learning Best Practices
• Lecture 7: Risk Mitigation Proposals for Language Models

Corrections or suggestions? Please file a GitHub issue.

---

Preliminary Materials

• Syllabus
• Basic Data Manipulation
• Primer on Technical Malpractice
• Whiteboard Notation

Lecture 1: Explainable Machine Learning Models

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

Lecture 1 Class Materials

• Lecture Notes
• Software Example
• Assignment 1:
  • Model evaluation notebook
  • Full evaluations results
• Reading: Machine Learning for High-Risk Applications, Chapter 2 (pp. 33 - 50) and Chapter 6 (pp. 189 - 217)
  • Check availablity through GWU Libraries access to O'Reilly Safari
• Lecture 1 Additional Materials

---

Lecture 2: Post-hoc Explanation

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

Lecture 2 Class Materials

• Lecture Notes
• Software Example
• Assignment 2
• Reading: Machine Learning for High-Risk Applications, Chapter 2 (pp. 50 - 80) and Chapter 6 (pp. 208 - 230)
  • Check availablity through GWU Libraries access to O'Reilly Safari
• Lecture 2 Additional Materials

---

Lecture 3: Bias Testing and Remediation

_{^{Source: Lecture 3 Notes}}

Lecture 3 Class Materials

• Lecture Notes
• Software Example
• Assignment 3
  • Model evaluation notebook
  • Full evaluations results
• Reading Machine Learning for High-Risk Applications, Chapter 4 and Chapter 10
  • Check availablity through GWU Libraries access to O'Reilly Safari
• Lecture 3 Additional Materials

---

Lecture 4: Machine Learning Security

_{^{Source: Responsible Machine Learning}}

Lecture 4 Class Materials

• Lecture Notes
• Software Examples:
  • Attacks for Red-teaming
  • Data Poisoning
  • Backdoor Attack
• Assignment 4
• Reading: Machine Learning for High-Risk Applications, Chapter 5 and Chapter 11
• Lecture 4 Additional Materials

---

Lecture 5: Machine Learning Model Debugging

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

Lecture 5 Class Materials

• Lecture Notes
• Software Examples:
  • Sensitivity Analysis:
    • Adversarial Example Search
    • Stress Testing
  • Residual Analysis
• Assignment 5
• Reading: Machine Learning for High-Risk Applications, Chapter 3 and Chapter 8
  • Check availablity through GWU Libraries access to O'Reilly Safari
• Lecture 5 Additional Materials

---

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)
• Reading: Machine Learning for High-Risk Applications, Chapter 1 and Chapter 12
• Lecture 6 Additional Materials

---

Lecture 7: Risk Mitigation Proposals for Language Models

_{^{A diagram for retrieval augmented generation. Source: Lecture 7 notes.}}

Lecture 7 Class Materials

• Lecture Notes
• Software Example
• Reading: Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile, pgs. 1-12, 47-53
• Lecture 7 Additional Materials

Example instructions for using class software in Google Colab

• Create a folder in your GWU Google Drive My Drive called DNSC_6330_Software

• To run the lecture examples:

  • Create a folder named Lecture_01 inside the DNSC_6330_Software
  • Save a copy of the class 01_Explainable_AI_Models.ipynb notebook into the Lecture_01 folder by using the File ->
    Save a Copy In Drive menu options, or download the class notebook and upload to your folder
  • In cell 1 of the 01_Explainable_AI_Models.ipynb update the path to the Lecture_01 folder:
    • likely %cd drive/My\ Drive/DNSC_6330_Software/Lecture_01/
    • Use the %cd and %ls commands to find your folder if needed
    • Generally the drive.mount('/content/drive/', force_remount=True) command can only be used once in a Colab session, so use %cd and %ls commands in a different cell or restart your colab session if you see strange errors
  • Download the example data from: https://drive.google.com/drive/folders/1jYZvT1j5khFnJC5NSqNeGiCNoOeib9YK?usp=sharing (Click triangle beside Data at top -> Download, then unzip and upload into DNSC_6330_Software folder)
  • Download some necessary Python code from: https://drive.google.com/drive/folders/1BPXxGp0QAKRl1ZP6Vd1xKuCwitiLyuy6?usp=sharing (Click triangle beside hrml_book at top -> Download, then unzip and upload into DNSC_6330_Software folder)
  • Upload these folders into the DNSC_6330_Software folder

• To run the homework templates:

  • Create a folder called assignments in the DNSC_6330_Software folder
  • Download then upload the notebook assign_1_template.ipynb into the assignments folder
  • Create a folder called data in the DNSC_6330_Software folder
  • Download the assignment data from the class GitHub: https://github.com/jphall663/GWU_rml/tree/master/assignments/data
  • Unzip the data files into CSV files and upload them into the data folder
  • In the assignment_1.ipynb add a cell before cell 1 that installs h2o and interpret: !pip install interpret h2o
  • In the assignment_1.ipynb add a cell before cell 3 that connects the notebook to the data:

from google.colab import drive
drive.mount('/content/drive/', force_remount=True)
# may need to be updated to location on your drive
%cd drive/My\ Drive/DNSC_6330_Software/assignments/
%ls

• Whenever asked, allow Colab to connect to your Google drive
• Delete any __pycache__ folders you see
• In the end the DNSC_6330_Software folder should look like:

DNSC_6330_Software
├── assignments
│   ├── assign_1_template.ipynb
├── data
│   ├── hmda_test_preprocessed.csv
│   └── hmda_train_preprocessed.csv
├── Data
│   ├── backdoor_testing
│   │   ├── constrained_backdoor_output.csv
│   │   ├── constrained_output.csv
│   │   ├── overfit_backdoor_output.csv
│   │   ├── overfit_output.csv
│   │   └── test_data.csv
│   ├── credit_line_increase.csv
│   ├── data_dictionary.csv
│   └── synthetic_data.csv
├── hrml_book
│   ├── explain.py
│   └── partial_dep_ice.ipynb
└── Lecture_01
    └── 01_Explainable_AI_Models.ipynb

• You can use the following commands in a colab notebook to check your file structure:

from google.colab import drive
drive.mount('/content/drive/', force_remount=True)
# may need to be updated to location on your drive
%cd drive/My\ Drive/DNSC_6330_Software/

!apt-get -y install tree

! tree

• Add new assignment templates into the assignments folder

• Add new code examples to Lecture_XX/XX_notebook_name.iypnb folders

  • For Lecture 2 that would be Lecture_02/02_Explainable_AI_Post_Hoc.ipynb

• For Lecture 7:

  • Click the Lecture 7 Software Example
  • Click the small triangle at the right of RAG_Example and click Download
  • Extract the downloaded zip file with a name like RAG_Example-20250428T160138Z-001.zip into a folder with a name like RAG_Example-20250428T160138Z-001
  • Inside the RAG_Example-20250428T160138Z-001 folder should be a folder named RAG_Example, upload the RAG_Example-20250428T160138Z-001/RAG_Example folder to the DNSC_6330_Software folder on your G-Drive
  • Click on the DNSC_6330_Software/RAG_Example/07_RAG_query.ipynb file and open it with Colab
  • To run the last session notebook (07_RAG_query.ipynb), you will need an access token from HuggingFace.co
    • Create a free Hugging Face account (https://huggingface.co/)
    • Generate an access token (Profile -> Access Tokens)
    • Preserve the text of the token somewhere secure, as you may not be able to see it again after it is granted
    • In the 07_RAG_query.ipynb notebook create a new secret (by clicking on the key icon on left) named HF_TOKEN and populated with the Hugging Face access token. Grant the notebook access to the secret by toggle the switch at left of the secret.
  • At the top of the notebook, update the path to your DNSC_6330_software folder, something similar to: %cd drive/My\ Drive/DNSC_6330_Software/RAG_Example
  • After adjusting these settings the notebook should run without errors