UCSD🔱 ECE285: Intro to Visual Learning (Spring 2025)

English Version | 中文版

Welcome to the official course repository for ECE285 – Intro to Visual Learning at UCSD, Spring 2025.
The repo collects all four programming assignments plus a creative final project covering classical CNNs, semantic segmentation, CAM visualization, and neural style transfer.

📚 Assignment 1 – Neural Networks in NumPy

• Build a two-layer fully-connected network entirely in NumPy
• Implement Linear, ReLU, Softmax, and Cross-Entropy modules
• Strictly vectorised (no explicit Python loops) to encourage clean math thinking
• Deliverables: 4 Python files + 2 notebooks + merged PDF submission

🖼️ Assignment 2 – Convolutional Neural Networks with PyTorch

• Gentle ramp-up from bare-bones tensors → nn.Module → nn.Sequential
• Train several CNNs on CIFAR-100; implement ResNet-10 from scratch
• Open-ended section: design any architecture & beat the baseline accuracy

🚦 Assignment 3 – Semantic Segmentation & Class Activation Maps

• Implement FCN-32s and FCN-8s, train both scratch & fine-tuned versions
• Add CAM to visualise class-specific regions
• Focus on reproducible training and qualitative mask inspection

🧪 Assignment 4 – Advanced Vision Challenge

• Explore a modern vision topic of your choice (e.g. ViT, object detection, self-supervised learning)
• Report findings and submit code + short write-up

🎨 Final Project – Neural Style Transfer

• Fork of the concise PyTorch Neural Style Transfer implementation
• Experiment with style/content weight trade-offs, TV loss, and different initialisations
• Provides ready-to-run scripts and Jupyter demos for artistic results