Detecting Unreliable Classifications from Quantized Models

Ananya Kotla, Jorge T. Gomez

Post-training INT8 quantization is typically validated by checking accuracy retention, which requires ground-truth labels often unavailable in low-resource clinical settings. We sh…