A multi-modal journey into the world of XAI — covering tabular, text, image data, and even transformers.
In late 2023, I found myself obsessing over a simple but powerful question:
"We trust models to make decisions — but do we understand how they do it?"
This question hit harder when I realized something troubling: even when traditional ML models give us a vague idea of what’s going on under the hood, LLMs (Large Language Models) remain mostly a black box. That tension between trust and transparency drove me into the rabbit hole of Explainable AI (XAI).
So in November 2023, fueled purely by curiosity, I began this project — not as a research requirement, not for a job, but to satisfy a need to know:
Can we really explain what our models are thinking?
This repo is a hands-on, multi-modal collection of XAI experiments using:
- SHAP (Shapley Additive Explanations)
- LIME (Local Interpretable Model-agnostic Explanations)
- BERTViz (Transformer attention visualization)
Each notebook dives deep into interpreting decisions made by models trained on tabular, text, or image data.
| Domain | Notebook(s) | Description |
|---|---|---|
| Tabular | XAI_SHAP.ipynbLIME_tabular_manual_interpretation.ipynbCovid19_XAI.ipynb |
Model explanations for structured data using SHAP and LIME. COVID case study included. |
| Text | XAI_LIME_text.ipynbLIME_text_manual_interpretation.ipynbBertViz.ipynb |
XAI for NLP tasks + BERT attention visualization. |
| Image | XAI_LIME_image.ipynb |
LIME applied to CNN predictions. |
| Concept Comparison | LIME_SHAP.ipynb |
Direct contrast between LIME and SHAP on the same problem. |
Most tutorials stick to one dataset, one domain, or one method. This project explores XAI across domains and includes:
- Manual interpretation walkthroughs
- Real-world data (e.g. COVID-19 clinical data)
- Transformer-level introspection with BERTViz
- A curiosity-first origin story — not just a checklist project