CIFAR-100 Image Classification Project

Overview

This project focuses on classifying images from the CIFAR-100 dataset using a custom Convolutional Neural Network (CNN). The project includes detailed analysis, model development, and performance tuning, resulting in achieving a Top 20 position (out of 200) on the Kaggle competition leaderboard.

Placement (20th place)

Dataset

• Dataset: CIFAR-100
• Images: 60,000 color images of 32x32 pixels
• Classes: 100 distinct categories grouped into 20 superclasses

Challenge

The goal was to classify images into their respective categories with high accuracy using deep learning techniques.

---

Features

1. Preprocessing & Data Augmentation

• Normalization: Standardized pixel values using channel-wise means and standard deviations.
• Data Augmentation:
  • Random horizontal flips
  • Random cropping with padding of 4 pixels

2. Model Design

• Architecture:
  • Three convolutional layers (32, 64, and 128 channels)
  • Max-pooling after every two convolutional layers
  • Dropout (50%) for regularization
  • Two fully connected layers
  • Output: 100 categories
• Optimizer: Stochastic Gradient Descent (SGD)
• Loss Function: Categorical Cross-Entropy
• Learning Rate: 0.05

3. Training Results

• Train Accuracy: 66.81%
• Test Accuracy: 56.03%
• Kaggle Leaderboard Score: 0.3184

---

Achievements

• Top 20 Leaderboard Position: Reached the top 20 in the Kaggle competition by balancing model simplicity and computational efficiency.
• Key Strategies for Success:
  • Leveraging data augmentation to improve generalization.
  • Employing dropout to reduce overfitting.
  • Optimizing training with an efficient CNN architecture.

---

Files

1. Report: Model Design and Performance Analysis

This document provides a detailed explanation of the model design, preprocessing methods, and performance analysis.

2. Notebook: ImageClassification_Kaggle_CNN.ipynb

This Jupyter Notebook contains the implementation of the CNN model, including:

• Data preprocessing
• Model architecture and training
• Performance evaluation and visualization

---

Instructions

1. Clone the repository:

   git clone <repository-url>

2. Open the Jupyter Notebook:

   jupyter notebook kaggle.ipynb

3. Follow the cells to understand the data preprocessing, model training, and evaluation process.

---

Lessons Learned

• The importance of balancing model complexity and computational efficiency.
• Regularization techniques, such as dropout, are essential for reducing overfitting.
• Leveraging data augmentation can significantly improve model robustness and generalization.

---

Future Work

• Experiment with more complex architectures like ResNet or DenseNet.
• Explore advanced regularization techniques (e.g., weight decay, mixup).
• Implement learning rate schedulers to optimize convergence.

---

Acknowledgments

This project was completed as part of COMPSCI 4ML3: Introduction to Machine Learning at McMaster University under the guidance of the course instructors.

---

Contact

Waleed Malik

• Kaggle: waleedmalik7
• Email: waleedmalik027@gmail.com