Seminar: Explainability in AI

Description

The idea of this seminar is to make the students familiar with basic algorithms in the field of artificial intelligence (AI). We will not cover neural networks, machine learning or deep learning as this is beyond the scope of this seminar. Alongside with basic AI we also want to introduce explainability. Explainable AI (XAI) is gaining more and more importance in recent years as AI systems take over in various fields. Two prominent examples are self-driving cars and medical diagnosis. Both require lots of confidence in AI and therefore it is of huge importance to be able to understand and comprehend decisions made by AI systems.

Other than in normal proseminars, you will get practical experience with the algorithms and implementations. You will implement one of the XAI algorithms for an AI algorithm (based on the sklearn pipeline) and present both together with the results in a 20 minutes presentation. Additionally, you will write a 5-pages report about your findings.

Despite the seminar officially is called "Explainability in AI for Drug Discovery" we will not discuss papers applying explainability techniques to AI models used to aid drug discovery. But we will dive deep into these explainability techniques to discuss and understand them. A good overview for XAI in drug discovery and this seminar is provided in Jiménez-Luna et al. published in Nature.

Requirements

The following requirements are not strictly enforced but will make your life way easier in the seminar. Basic knowledge in:

• python coding,
• Jupyter notebooks,
• sklearn, and
• Elements of [Statistical|Machine] Learning

What Do You Need To Do?

• Read and present one of the seminar papers (as assigned in the kick-off meeting).
• The presentations are 20 minutes with additional 10 minutes of questions from the audience.
• Implement your XAI method in a Jupyter notebook to get hands-on experience (no worries, you are can use libraries, etc.)
• Write a 5-pages report on your findings.
• You need to attend all talks.

How Are the Grades Computed?

Presentation (30%)

Assessed on clarity, depth of understanding, and the quality of your answers to audience questions.

Notebook (30%)

We will check for:

• Completeness
• Documentation
• Executability (can I run the notebook from top to bottom without raising errors?)

Report (30%)

The report should include:

• A short explanation of how the model you explained works
• An explanation of how the XAI technique works. What are its (dis-)advantages?
• Comparison to other methods in this seminar or beyond
• Your own thoughts on this method

Participation (20%)

Active engagement during other students' presentations. Ask questions — it counts toward your grade and improves the seminar for everyone.

Plagiarism

Registration

Please register to this seminar by writing an email to Roman Joeres before 30.04.2023 23:59. Please also attach your transcript of records which can be downloaded from the LSF/HISPOS.

Organisational Details

Important Dates

Topics

All notebooks will deal with the Breast Cancer Dataset from the sklearn package to compare the algorithms performance and have interpretable features and comparable features. We will mainly rely on the sklearn python library. Topics 1. and 2. will explained based on their implicitly computed feature importance. Topics 3. and 4. apply their XAI technique to a linear regression model. From topic 5 on, you will explain a neural network as true black box.

1. Decision Trees and Random Forrest with Gini Index Assigned to: Jis Kochuniravathu Saji
   sklearn docu sklearn DT sklearn RF
2. Support Vector Classifier with Linear Kernel Assigned to: Kriti Maurya
   ESL II pp. 417 sklearn
3. SHAP Assigned to: Shilpa Sharma
   Paper PyPI GitHub Notebook
4. LIME Assigned to: Annmariya Elayanithottathil Sebastian
   Paper PyPI Notebook
5. Permutation Feature Importance Assigned to: Evgenia Khodzhaeva
   Paper sklearn Notebook
6. DiCE Assigned to: Jibin Varghese
   DOI PyPI GitHub
7. NiCE Assigned to: Muhammad Hamid
   Paper PyPI GitHub Notebook

Additional reading, papers that help understanding the bigger picture:

• The Elements of Statistical Learning, 2nd Edition
• Interpretable Machine Learning
• XAI Tutorials
• Counterfactuals