Revealing Similar Semantics Inside CNNs: An Interpretable Concept-based Comparison of Feature Spaces

CNNs learn similar concepts across different architectures: The research demonstrates that different CNN designs learn comparable semantic information despite their structural differences, suggesting that fundamental visual concepts (like edges, textures, and objects) emerge naturally during training regardless of how the network is built.

Semantic concepts appear at predictable depths: The study found that similar concepts emerge at similar relative layer depths across networks, meaning a concept learned in the middle layers of one architecture tends to appear at a proportionally similar depth in another—this consistency could help predict where to look for specific learned features.

Current performance metrics miss important differences: The paper highlights that comparing CNNs based only on accuracy scores overlooks crucial information about how networks internally represent knowledge, which is particularly important for safety-critical applications where understanding the reasoning process matters as much as the final answer.

New tools enable "opening the black box": By using explainable AI techniques to extract and compare concept activation vectors, the researchers provide a practical method for examining what's happening inside neural networks—moving beyond treating them as opaque systems and toward genuine transparency.

Interpretability is essential for trustworthy AI: The work underscores that safety-critical applications (like medical imaging or autonomous vehicles) require more than just good performance; they need transparency about what the AI has learned and how it makes decisions.