Pruning coupled with learning, ensembles of minimal neural networks, and future of XAI

Neural network pruning can improve explainability: The paper demonstrates that removing unnecessary connections from neural networks while maintaining performance creates "logically transparent" networks that are easier to interpret and explain their decision-making processes.

Multiple valid solutions exist for the same problem: The research found that eight different minimal neural networks could be created for the same forecasting task (US presidential election prediction), highlighting that data-driven AI doesn't produce a single "correct" answer—an important consideration for responsible deployment of AI systems.

Ensemble approaches can increase reliability and transparency: By combining multiple minimal neural networks as a "committee" that votes on decisions, the approach creates systems where each network's reasoning can be individually examined and verbalized, potentially reducing bias and improving accountability.

Pruning serves multiple responsible AI goals simultaneously: The paper shows that network pruning addresses several key concerns in responsible AI including preventing overfitting, reducing computational resource requirements, and enabling explainability—making it a practical tool for more sustainable and interpretable AI systems.

The principle "what controls training should control pruning" offers a systematic framework: Using gradient-based sensitivity measures to guide pruning creates a principled, mathematically grounded approach to simplifying networks, which could help standardize explainability efforts across different AI applications.