Learning XAI: Explainable Artificial Intelligence

Course details

This course focuses on the data-based and mathematical factors contributing to bias in generative AI. Join instructor Jazmia Henry as she explores how data curation, analytical techniques, and post-training constraints can mitigate harmful biases embedded in these models. Through real-world examples and case studies, the course aims to help you understand and apply strategies for creating fairer, more transparent AI systems.

Learning objectives
Understand the data sources and mathematical assumptions behind generative AI that contribute to biased outcomes.
Identify and evaluate biases in LLMs using data curation, statistical analysis, and counterfactuals.
Implement practical techniques to constrain models post-deployment to reduce the risk of bias reintroduction.
Apply parity metrics to measure and adjust bias levels in model outputs.
Explore case studies on biased language models, assessing failures and opportunities for improved training methods.

Concepts

0. Introduction

01 - XAI - the key to fair and transparent AI for all

1. Understanding the AI Landscape

02 - Bias in AI and its impact
03 - How does generative AI work
04 - Challenges of bias mitigation

2. Bias in GenAI

05 - Bias in data - Bad data in, bad data out
06 - Prompt engineering crash course
07 - Bias in mathetmical assumptions

3. Bias in Model Training

08 - Introduction to GenAI model training
09 - Demographic parity and its applications
10 - Equal opportunity parity for evaluating fairness
11 - Equalized odds parity to compare subgroup performance
12 - HELM
13 - Red-teaming
14 - Exercise - Building an evaluation pipeline

4. Counterfactuals in Bias Analysis

15 - What is a counterfactual
16 - Concept of counterfactuals and what-if scenarios
17 - Applications of counterfactuals for transparency
18 - Counterfactual modeling for improved fairness
19 - Exercise - TensorFlow counterfactual

5. Data Sampling for Bias Reduction

20 - The power of sampling data
21 - Data validation and evaluation
22 - Overview of stratified sampling
23 - Sampling based on demographics
24 - Random sampling within stratified groups
25 - Testing model performance across diverse samples
26 - Exercise - Design sampling criteria with RAI principles

6. Post-Training Model Constraints

27 - Introduction to model constraints to control bias
28 - Power of flagging, pruning, and nudging
29 - Using human-in-the-loop tools
30 - Exercise - Model constraint pipeline

7. Case Study - Breakdown of a GenAI Story

31 - Explore an example of GenAI output

Conclusion

32 - Continuing your XAI learning journey