Efficiently Transforming Neural Networks into Decision Trees: A Path to Ground Truth Explanations with RENTT

Decision trees can provide truly faithful explanations of neural networks: Unlike many existing explainability methods that may misrepresent how a neural network actually works, the RENTT algorithm creates exact equivalent decision tree representations that guarantee the explanations match the network's actual decision logic.

The new RENTT algorithm makes tree conversion practical at scale: Previous attempts to convert neural networks to decision trees struggled with computational efficiency and couldn't handle larger networks. RENTT overcomes these limitations by being both runtime and memory efficient, making the approach viable for real-world applications.

The method works across different types of neural networks: The research extends beyond simple networks to handle convolutional neural networks (used in image recognition) and recurrent neural networks (used in sequential data), making the approach more broadly applicable to modern AI systems.

Ground truth feature importance becomes calculable: By converting networks to decision trees, researchers can now determine which features (inputs) a neural network actually relies on for its decisions—at the global model level, for specific regions of data, or for individual predictions—providing concrete answers rather than approximations.

Interpretability remains a challenge even with exact explanations: While RENTT solves the faithfulness problem, the resulting decision trees can still be too complex for humans to easily understand, indicating that converting to decision trees is a necessary but not sufficient step toward truly trustworthy AI systems.