Compressing Deep Neural Networks Using Explainable AI

Explainability can guide smarter compression decisions: Rather than randomly removing parts of neural networks, this research uses explainable AI (specifically Layer-wise Relevance Propagation) to identify which components actually matter for the model's decisions, enabling more targeted and effective compression.

Significant efficiency gains with minimal accuracy trade-offs: The proposed method achieved a 64% reduction in model size while improving accuracy by 42% compared to previous explainability-based compression approaches, suggesting that understanding why a model works helps preserve its performance during compression.

Making AI more deployable on everyday devices: By reducing memory requirements through intelligent compression guided by explainability techniques, this work makes it more feasible to run sophisticated AI models on resource-limited devices like smartphones and IoT sensors, expanding where AI can be responsibly deployed.

Transparency and efficiency can work together: This research demonstrates that the goals of responsible AI—making models more interpretable and explainable—can actually align with practical engineering goals like reducing computational costs, rather than being in conflict with them.

Different parts of neural networks deserve different treatment: The method applies different levels of compression (quantization) to different parameters based on their importance scores, reflecting the insight that not all components of a neural network contribute equally to its decisions.