Concept Bottleneck Network

[max-acc]

This issue serves as a note to explore Concept Bottleneck Networks as a potential future direction for the Particle Track reconstruction pipeline.

CBM introduce an intermediate layer that predicts a set of human interpretable concepts, which serve as the basis for the final task prediction.

Note

There has to be done more research on how this might be possible to implement.

Potential Benefits

• Improved interpretability of model's decisions, allowing to inspect physics motivated concepts.
• Better alignment with domain knowledge.
• Easier debugging.