Explainable Artificial Intelligence (VTR UGE 21), day 1 of 5

Foundations of Interpretability

Prof. Dr. Utku Kose, Süleyman Demirel University

Glass-box lab

Part A compares interpretable models on simulated tumours with two measurements and records accuracy against the number of values a reader must hold in mind. Part B places eight explanations from practice on the map of methods. Part C follows five patients through a small decision tree by hand.

Part A: Accuracy against readability

Choose a model family and its setting, then press Run and add to scoreboard. The panel shows the decision regions on held-out tumours. Every run is added to the scoreboard below as one point: Test accuracy against the number of values a reader must hold in mind to reproduce a prediction.

Model family

Scoreboard

Find the run with the highest accuracy for the fewest values. Then increase the depth until the accuracy stops improving: The point where the frontier flattens is the place to stop adding complexity.

Run	Model	Values to read	Test accuracy

Continue in Colab, section 5: the same comparison with all 30 measurements and a random forest.

Part B: Sort the explanations

Each card describes an explanation used in practice. Decide whether it describes the whole model or one prediction, and whether it comes from a model that is readable by design or is computed after training. Then press Check the cards.

Continue in Colab, section 7: read the models of your own field.

Part C: Read a tree by hand

The tree below is a small illustrative classifier for tumours. Follow each patient from the top question to a leaf, choose the prediction and press Check the paths. The feedback shows the path the tree takes.

Is worst concave points at most 0.14?
├─ yes: Is worst radius at most 17.0?
│       ├─ yes: benign
│       └─ no:  malignant
└─ no:  Is worst texture at most 19.0?
        ├─ yes: benign
        └─ no:  malignant

Patient	Worst concave points	Worst radius	Worst texture	Your prediction	Feedback

Continue in Colab, section 4: draw the full tree and follow any patient.