Machine Learning Project: CIFAR-3 Classification

This project focuses on image classification using various machine learning techniques applied to the CIFAR-3 dataset, a subset of CIFAR-10. The dataset includes grayscale images of three classes: Automobile, Deer, and Horse. The project explores methods from dimensionality reduction to deep learning models, with the goal of developing effective classifiers.

Project Overview

This repository contains:

• Python code implementing several machine learning models to classify images in the CIFAR-3 dataset.
• Data preparation and visualization techniques to better understand the dataset.
• Application of PCA for dimensionality reduction, traditional machine learning classifiers, and deep learning models including an MLP and a CNN.

Dataset

The CIFAR-3 dataset consists of:

• X_cifar_grayscale.npy: Grayscale images, a tensor of shape (18000, 32, 32).
• Y_cifar.npy: Labels for each image, with classes represented as integers (0=Automobile, 1=Deer, 2=Horse).

Requirements

• Python with dependencies: numpy, matplotlib, scikit-learn, and tensorflow.
• CIFAR-3 dataset files in the working directory.

Code Description

Data Loading and Preprocessing

• Data loading: The dataset is loaded and prepared for use in various models.
• Preprocessing: The grayscale images are flattened to a vector form (for PCA) or reshaped for CNN processing.

Principal Component Analysis (PCA)

• PCA Dimensionality Reduction: Applied on grayscale images, with experiments using various numbers of components to understand data variance and feature reduction.
• Reconstruction Visualization: Reconstructed images from PCA to demonstrate data loss at different component levels.

Supervised Machine Learning Models

• Logistic Regression and Gaussian Naïve Bayes: Basic classifiers trained on the grayscale data, with and without PCA-based dimensionality reduction.
• Evaluation: Accuracy scores are calculated on training and test sets.

Deep Learning Models

• Multilayer Perceptron (MLP): A fully connected neural network with hidden layers, dropout, and L2 regularization. Trained on PCA-reduced data.
• Convolutional Neural Network (CNN): A convolutional model to handle the spatial structure in color images, with regularization and batch normalization for robust training.

Training and Evaluation

Each model is evaluated on:

• Training Accuracy and Test Accuracy: Metrics are tracked to evaluate model performance.
• Loss and Accuracy Plots: Visualizations of training progress, with validation metrics included.

How to Run the Code

1. Install the dependencies listed in requirements.txt.
2. Place dataset files (X_cifar_grayscale.npy, and Y_cifar.npy) in the working directory.
3. Run the script:

   python script.py