Tuesday, 15 December 2026
12:45 Join in
13:00 Welcome, Motivation & Introduction to the Kandinsky Pattern Problem
13:15 How do humans solve visual reasoning problems?
- Introduction to Kandinsky Patterns
- Reasoning from a small number of labelled examples
- Progressively increasing pattern complexity
- Collaborative exercise: participants infer and discuss the hidden classification rules in small groups
14:15 Break
14:30 How does the model solve the same problem?
- Training a lightweight neural network on Kandinsky Patterns
- Inspecting predictions, errors and potential shortcuts
- Applying DeepLIFT/Grad-CAM, LIME and SHAP
- Hands-on lab: generate explanations and compare them with the participants' original hypotheses
15:30 Break
15:45 Human reasoning vs. model explanations
- Moving to more complex Kandinsky patterns
- Do explanations reveal the concepts we expected the model to learn?
- Plausible explanations vs. faithful explanations
- Collaborative challenge & discussion: human hypothesis → model prediction → explanation → interpretation
17:00 End of course