A Multi-Method Interpretability Framework for Probing Cognitive Processing in Deep Neural Networks across Vision and Biomedical Domains

Harshini Suresha, Kavitha S H

Keywords: Reliable AI, Model Calibration, Interpretability Fidelity, Attribution Robustness, Cross-Domain Generalization, Trustworthy Machine Learning

TL;DR: We propose TriggerNet, a triggered interpretability framework combining Grad-CAM, RISE, FullGrad, and TCAV to yield selective, reliable explanations.

Abstract: Interpretable deep learning remains a central challenge across high-stakes domains such as agriculture, healthcare, and vision-based diagnostics. We present TriggerNet, a novel framework that integrates Grad-CAM, RISE, FullGrad, and TCAV to generate class-discriminative, high-fidelity explanations. TriggerNet is evaluated on three diverse datasets: (i) Red Palm Mite-affected plants (11 species, 4 disease stages), (ii) MedMNIST (PathMNIST, OrganMNIST), and (iii) CIFAR-10. Our framework leverages CNNs, EfficientNet, MobileNet, Vision Transformers, and ResNet50, combined with Snorkel-based supervision. Quantitatively, TriggerNet achieves accuracies of 97.3\% (plants), 94.2\% (PathMNIST), and 92.8\% (CIFAR-10), while improving interpretability with a 21.4\% p-score gain and 16.7\% lower Brier score. Qualitatively, TriggerNet produces focused, meaningful visual explanations as it aligns with anatomical features in medical scans, localizes plant symptoms like yellowing and webbing with near-human accuracy, and highlights object boundaries over background noise in CIFAR-10.

Submission Number: 47