IMPACTX: improving model performance by appropriately constraining the training with teacher explanations

Explainability can improve AI performance, not just transparency: IMPACTX demonstrates that using explainability techniques during training—rather than only after deployment—actually makes AI models perform better. The approach improved accuracy across all three tested deep learning models (EfficientNet-B2, MobileNet, and LeNet-5) on standard image datasets.

AI systems can learn to focus on the right features: By incorporating explanation-based attention mechanisms during training, the model learns to prioritize relevant features for decision-making. This means the AI becomes more focused and reliable without requiring humans to manually guide it toward important information.

Built-in explanations reduce need for external tools: IMPACTX provides explanations directly as part of the model itself, eliminating the need to use separate explainability tools after the model makes decisions. This makes responsible AI deployment more practical and efficient in real-world applications.

Consistent improvements across different scenarios: The method was tested on multiple popular deep learning architectures and datasets (CIFAR-10, CIFAR-100, and STL-10), showing that the benefits aren't limited to one specific use case—suggesting broader applicability for improving AI systems responsibly.

Automated approach requires no human intervention: Unlike some explainability methods that depend on human feedback or external knowledge, IMPACTX works automatically during training, making it easier to implement responsible AI practices without adding significant manual oversight requirements.