sarib = {
"degree" : "BS Artificial Intelligence — 3rd Year",
"location" : "Islamabad, Pakistan",
"research" : ["Model Interpretability", "Calibration", "Computer Vision", "NLP"],
"belief" : "Accuracy tells you what a model gets right.",
"mission" : "Calibration and interpretability tell you whether to trust it.",
}I study the gap between what AI models claim and what they actually deliver. My projects go beyond training and evaluation — I investigate why models predict, where they look, and whether their confidence can be trusted.
|
Post-hoc calibration analysis of ResNet-50 on CIFAR-10. Measured systematic overconfidence using ECE and MCE. Applied temperature scaling — a single scalar T — to reduce ECE by 67% without retraining. Key finding: High top-1 accuracy and good calibration are independent properties. Most models only optimise for one.
Based on Guo et al. (2017), ICML |
Comparative saliency analysis across ResNet-50, EfficientNet-B0, and ViT-B/16 on ImageNette. Implemented Grad-CAM for CNNs and Grad-CAM with patch reshape transform for ViT. Key finding: CNNs focus on local discriminative patches. ViT attends globally — even when all three models agree on the correct label.
Based on Selvaraju et al. (2017), ICCV |
|
Sequence classification using bidirectional LSTM with attention for multilingual fake news detection.
|
Language model-based autocomplete system built on neural sequence modeling approaches.
|
Methods & Techniques:
- Grad-CAM · Attention Rollout · Saliency Mapping
- Temperature Scaling · ECE/MCE · Reliability Diagrams
- BiLSTM · Multi-head Attention · Transformer Architectures
- ResNet · EfficientNet · Vision Transformer (ViT)
- 📖 Reading papers in model calibration, interpretability, and vision transformers
- 🔬 Building projects that study model trustworthiness beyond accuracy
- 🎯 Goal: AI research that makes deployed models reliable and explainable
