Neural-Networks-from-Scratch

About:

This project provides a hands-on implementation of fundamental neural network architectures using PyTorch. It covers:

• Linear and Logistic Regression: Demonstrates how these classic algorithms can be represented as single-layer neural networks.
• Multilayer Perceptron (MLP): Builds a 3-layer MLP for image classification on the MNIST dataset.
• Convolutional Neural Networks (CNNs):
  • Implements the 2D convolution operation from scratch.
  • Shows how to learn a convolutional kernel for edge detection.
  • Implements the ResNet18 architecture for image classification.

This project is ideal for anyone learning about neural networks and wanting to gain a deeper understanding of their inner workings.

Usage:

1. Synthetic Data Regression: Run the code in section 1.1 to generate synthetic data and train a linear regression model.
2. MNIST Classification:
   • Run the code in section 1.2 to train a single-layer neural network for MNIST classification.
   • Run the code in section 2 to train a 3-layer MLP for MNIST classification.
3. Convolutional Neural Networks:
   • Run the code in section 3.1 to experiment with the corr2d function and understand the convolution operation.
   • Run the code in section 3.2 to learn a kernel for edge detection.
   • Run the code in section 3.3 to train the ResNet18 model on the MNIST dataset.

Key Features:

• From Scratch Implementation: Provides a deeper understanding by implementing core components manually.
• Multiple Architectures: Covers a range of neural network architectures, from basic to advanced.
• Clear Explanations: Includes comments and explanations within the code to guide understanding.
• Visualization: Uses matplotlib to visualize training progress and results.

Contributor: AISHWARYA NAYAK (Contributions are welcome! Feel free to open issues or submit pull requests.)