Project on exploring how different vision models “see” and analyze the images. We compare ConvNeXt, DeiT, and MLP-Mixer using Grad-CAM, attention maps, and saliency, with tools to run inference, analyze results, and visualize model focus regions.

[GCPR 2023] Zero-shot Translation of Attention Patterns in VQA Models to Natural Language

PathAI: An integrated AI healthcare platform featuring ViT-based skin cancer screening, a Google HeAR ensemble for acoustic respiratory analysis, and Luna, an AI-powered menstrual cycle health advisor.

Interpretability analysis of Vision Transformers using Attention Rollout, Gradient Rollout, and a novel FAF/BAF metric with Mask R-CNN. Includes experiments on DeiT-Tiny comparing CNN and ViT explanation behavior, showing ~25% improvement in semantic focus using gradient-based rollout.

Vision Transformer (ViT) for image classification using self-supervised learning (MAE) with Explainable AI (Grad-CAM & Attention Rollout).

Automated rubric-anchored response scorer with fairness audit and explainability — QWK 0.931, adverse-impact detection, attention-rollout highlights

Explainable deep learning for multi-label thoracic disease classification on NIH ChestX-ray14. Benchmarks 7 architectures (CNN, ResNet50, DenseNet121, Xception, ViT, ConvNeXt) with Grad-CAM, SHAP, LIME, and Attention Rollout.