si4onnx: A Python package for Selective Inference in Deep Learning Models

Statistical rigor improves AI reliability: The package addresses a critical gap where common explanation methods like CAM (Class Activation Mapping) and VAE-based anomaly detection identify important image regions, but without statistical testing to confirm these regions are genuinely meaningful rather than artifacts.

Type I error control prevents false conclusions: si4onnx implements hypothesis testing with controlled type I error rates, meaning it reduces the risk of incorrectly claiming an AI system has identified something significant when it hasn't—a key requirement for trustworthy AI systems.

Bridges explanation and validation: While XAI (Explainable AI) techniques show what a model is looking at, selective inference adds the crucial step of verifying those explanations are statistically sound, addressing a major gap between interpretability and reliability.

Practical accessibility across frameworks: By supporting PyTorch and TensorFlow models through the ONNX format, the package makes statistical validation techniques available to practitioners using mainstream deep learning tools, lowering barriers to responsible AI implementation.

Applicable to high-stakes domains: The focus on anomaly detection and image analysis suggests relevance for critical applications (medical imaging, security systems) where false positives from unreliable explanations could have serious consequences.