Harshini Suresha

A Multi-Method Interpretability Framework for Probing Cognitive Processing in Deep Neural Networks

Multi-Method Model Interpretability

Year2025
LocationNeurIPS 2025 workshops
TagsGrad-CAM, TCAV, ViT, MedMNIST

TriggerNet extended across vision, agricultural and biomedical imaging (MedMNIST, CIFAR-10): 97.3% on plant images, a 21.4% gain in p-score and fidelity measured by Pointing Game, Deletion/Insertion AUC and IoU.

Jan 2025 – Jun 2025 at NeurIPS 2025 workshops.

Workflow figureOpen full size ↗
Figure for A Multi-Method Interpretability Framework for Probing Cognitive Processing in Deep Neural Networks
Methods
  1. 01
    Datasets

    Red palm mite plant images (11 species, four disease stages), MedMNIST PathMNIST and OrganMNIST, and CIFAR-10.

    MedMNIST, CIFAR-10
  2. 02
    Models

    Trained and compared several architectures on each domain.

    CNN, EfficientNet, MobileNet, Vision Transformer, ResNet50
  3. 03
    Multi-method explanations

    TriggerNet fuses complementary attribution methods into class-discriminative explanations of what the model is looking at.

    Grad-CAM, RISE, FullGrad, TCAV
  4. 04
    Reliability

    Weak supervision for annotation, and explanation quality scored with p-score and Brier score alongside accuracy.

    Snorkel, p-score, Brier score
  5. 05
    Faithfulness

    Measured whether explanations point at the right thing against expert annotations.

    Pointing Game, Deletion / Insertion AUC, IoU
  6. 06
    Results

    97.3% on plant images, 94.2% on PathMNIST, 92.8% on CIFAR-10, with a 21.4% better p-score and 16.7% lower Brier score than baselines.

    Python, PyTorch / TensorFlow
← Back