Neural Network From Scratch

This project implements a neural network using Python. The implementation includes key features such as model architecture, training, evaluation, and optimization. The project is structured to be modular and easily extensible.

Table of Contents

• Features
• Implementation
  • Model Definition
  • Training Process
  • Visualization
• Installation
• Usage
• License

Features

• Customizable Model Architecture: Supports multiple layers with configurable activation functions and layer sizes.
• Training and Evaluation: Implements training loops with backpropagation and optimizers.
• Dataset: Uses the EMNIST dataset. Loads and preprocesses datasets efficiently.
• Performance Metrics: Computes accuracy, precision, recall, and F1 score for classification tasks.
• Hyperparameter Tuning: Allows modification of learning rates, batch sizes, and optimization techniques.
• Report: Generates a detailed report with training and evaluation metrics.

Implementation

The core implementation includes the following components:

Model Definition

It consists of:

• Input Layer
• Hidden Dense Layers (with ReLU as activation function)
• Dropout Layer (for regularization)
• Output Layer (for classification or regression tasks)

Training Process

The training process involves:

1. Forward propagation
2. Loss computation (cross-entropy for classification)
3. Backpropagation
4. Parameter updates using ADAM optimizer
5. Performance tracking

Visualization

The implementation includes visualization tools to analyze:

• Training and validation loss
• Accuracy trends
• Model performance metrics

Installation

To run this project, ensure you have the following dependencies installed:

pip install -r requirements.txt

Usage

Run the neural_net.ipynb notebook step by step to train and evaluate the neural network.

License

This project is licensed under the MIT License.