Author: Mubeen Ahmed
Roll No: 25280101 (MS-AI)
Course: AI 600 - Deep Learning, Spring 2026
Date: April 8, 2026
This repository contains the full solution to Assignment 3 for the AI 600 Deep Learning course. The project is divided into an analytical study of CNN mechanics and a series of programming experiments focusing on model design, shortcut learning, and interpretability.
- Toy-CNN Derivation: Manual forward and backward pass calculations for a convolutional layer, including ReLU activation gradients.
- Filter Gradient Analysis: Identification of dominant filter weights based on input pixel intensity and upstream gradients.
- Pooling & Normalization: Mathematical analysis of Global Max Pooling, Layer Normalization, and the spatial properties of translation invariance.
- Custom MNISTNet: Design of a compact CNN with fewer than 50,000 parameters (final count: 29,882) achieving a test accuracy of 99.27%.
- Shortcut Learning (Colored-MNIST): Investigation into how models prioritize simple color cues over complex shape features, resulting in an 8.55 percentage point accuracy drop on unbiased data.
- Transfer Learning (STL-10): Fine-tuning a ResNet-18 model with a frozen ImageNet backbone to achieve 76.75% accuracy on a limited dataset.
- Interpretability: Implementation of GradCAM to visualize class-discriminative heatmaps and analyze failure modes like background confusion.
The complete code and project files can be accessed via this GitHub repository.
└── Output/ # Generated results and visualizations
├── 25280101_deep_learning_PA3.ipynb # Implementation of MNIST, C-MNIST, and STL-10 experiments
├── 25280101_mubeen_ahmed_deep_learning_assignment_3_report.pdf # Technical report containing derivations and analysis
├── README.md # This file