Building Neural Networks for Pixel Data

This project demonstrates how to build and optimize neural networks for image classification using the Fashion MNIST dataset. Rather than manually extracting features (like with PCA), this demo shows how neural networks can automatically learn hierarchical features from raw pixel data.

Overview

This notebook covers:

• Building baseline neural networks with Keras/TensorFlow
• Experimenting with network architecture (depth and width)
• Visualizing training progress with accuracy and loss curves
• Implementing early stopping to prevent overfitting
• Comparing different network configurations

The Fashion MNIST on Kaggle contains 70,000 grayscale images of clothing items (28×28 pixels = 784 features per image), making it ideal for demonstrating neural network capabilities.

Setup

Prerequisites

• Python 3.x
• VS Code with Jupyter extension
• Kaggle account (for dataset download)

Installation

1. Clone or download this repository

2. Download the Fashion MNIST dataset

   • Visit Fashion MNIST on Kaggle
   • Download fashion-mnist_test.csv and fashion-mnist_train.csv
   • Place both files in the data/ directory

3. Install required packages

   Open a terminal in VS Code and run:

   pip install numpy pandas matplotlib tensorflow scikit-learn

Running the Notebook in VS Code

1. Open the notebook

   • Open pixel-neural-network.ipynb in VS Code
   • VS Code will automatically activate the Jupyter notebook interface

2. Select a Python kernel

   • Click on "Select Kernel" in the top right corner
   • Choose your Python environment (preferably one with the required packages installed)

3. Run the cells

   • You can run cells individually by clicking the play button next to each cell
   • Or run all cells sequentially using "Run All" from the toolbar
   • The notebook is designed to run from top to bottom

Project Structure

DSMII-Clothing-Neural-Network/
├── pixel-neural-network.ipynb    # Main notebook with neural network experiments
├── data/
│   ├── fashion-mnist_train.csv   # Training dataset (download from Kaggle)
│   └── fashion-mnist_test.csv    # Test dataset (download from Kaggle)
└── README.md                      # This file

Notes

• The notebook uses verbose=0 for most training runs to reduce output clutter
• Early stopping is demonstrated to show how to prevent overfitting
• All models use the Adam optimizer and sparse categorical crossentropy loss
• A 20% validation split is used to monitor model performance during training