CNN from Scratch: A Deep Dive into Neural Network Fundamentals

This repository contains the implementation of Convolutional Neural Networks (CNNs) built from scratch using minimal libraries like NumPy, CuPy, and Numba. This project focuses on understanding the foundational operations and components of neural networks, including custom CUDA kernels and detailed feature visualizations.

Features

• Custom Layers & Operations: Fully connected, convolution, max pooling, and flatten layers implemented from scratch.
• CUDA Accelerated: Custom kernels for efficient GPU computations.
• Dataset Benchmarks: Models trained on:
  • GTSRB (Traffic Sign Classification)
  • CIFAR-10 (Object Recognition)
  • Fashion-MNIST (Clothing Classification)
• Feature Visualization: Plots of activation maps, gradients, and learned kernels.

Don't miss out on the visualizations!