Welcome to the Interpretable Artificial Intelligence course repository offered at the University of Tehran. This repository contains code for assignments and projects completed during the course. The course by:
1- Robustness and Generalization:
- General AI
- Bias-Variance trade-off and regularization
- Inductive biases vs. generalization
- Thought process to choose inductive biases
- How to use inductive biases in a model
- Self-supervision and its useful inductive biases
- Multi-task learning and its emerging properties
- The emergence of deduction (AI) from induction (learning)
- Learning with loss function vs learning with rules
- Generalization vs robustness
- Generalization to different domains
- Robustness to input noise in a given domain
- Trade-offs between robustness and generalization
- Learning with noisy inputs
- Effects of noise on generalization
2- Interpretability and Explainability:
- Prescriptive models (models that make decisions for us)
- Why predictive models need some causal understanding
- How prescriptive models make decisions affecting humans
- Tragedies in prescriptive machine learning
- Why interpretability matters in prescriptive models
- Interpretable models
- Linear models, decision trees, etc. and how to interpret them.
- The relation between model interpretability vs causal analysis
- Trade-off between interpretability and prediction accuracy
- Practical cases where interpretability becomes more important than accuracy
- Explaining the workings of machine learning models
- Explainability vs interpretability
- Importance of explanation in advancing machine learning
- Model-dependent vs model-agnostic methods
- Example-based methods
- Global and local explanation techniques
- Understanding the workings of deep models
3- Safety and Security:
- Security attacks in machine learning:
- poisoning of a dataset
- backdoor attacks and trojan attacks
- open-access models vs closed-access models
- Security attacks in generative models and language models
- Watermarking
- Defense against attacks:
- Trojan detection and characterization
- Input/output monitoring
- Safety and reliability:
- Supervised/unsupervised Anomaly detection
- Out of distribution detection
- Accumulation of error in autoregressive models
- Safety in prescriptive models
- How to quantify reliability (domain shift, augmentation, adversarial)
4- Fairness, Transparency, Ethics and Privacy:
- Fairness and justice
- How to define fairness in mathematical terms
- Controversies in expressing defining justice mathematically
- How to measure fairness/bias
- Use of interpretable models to analyze fairness
- Ways to enforce fairness in machine learning models
- Transparency and Opacity in decision making
- In what contexts decision making process should be transparent
- Importance of interpretability in transparency
- Transparency regulations
- Ethics of AI
- How AI affects the job market, democracy, environment
- Who is legally liable if an AI causes an injury
- The use of chatbots and their impact.
- Fair use of data
- Regulatory bodies of AI
Please find below a brief overview of the contents of this repository:
-
HW1/: In this directory, you'll find the code for Assignment 1, which centers around training a robust model through data augmentation and exploring the use of the Angular loss function. Additionally, we examine the model's resilience against fast gradient methods and its performance when subjected to noisy data. -
HW2/: In this directory, you'll find the code for Assignment 2, which delves into various interpretability methods like SHAP, knowledge distillation, using D-rise to generate saliency maps, and employing LIME to interpret MobileNet's decisions. -
HW3/: This directory contains code for Assignment 3, which centers around backdoor attacks and out-of-distribution detection.
This repository is for archival and reference purposes only. The code here might not be updated or maintained. Use it at your own discretion.