Deep Learning Mini-Project: Evaluation and Optimization of Modified ResNet Architecture on CIFAR-10

Overview

This project focuses on creating and evaluating a modified ResNet model for image classification using the CIFAR-10 dataset. The project involves optimizing the ResNet architecture to improve accuracy while maintaining computational efficiency. This repository contains the source code, model checkpoints, and training scripts.

Table of Contents

• Introduction
• Results
• Future Work
• Contributing

Introduction

Image classification is a critical task in computer vision with applications ranging from facial recognition to autonomous vehicles. In this project, we build on the ResNet architecture and introduce modifications to enhance its performance on the CIFAR-10 dataset. The CIFAR-10 dataset consists of 60,000 32x32 color images across 10 classes. By modifying the ResNet architecture, we aim to optimize computational resources while achieving high accuracy.

Results

After training the modified ResNet model for 30 epochs, the model achieved a training accuracy of 85.49% and a test accuracy of 79.33% on the CIFAR-10 dataset. The steady increase in accuracy and decrease in loss across epochs indicates effective learning and model convergence.

Future Work

Potential directions for future research and improvements include:

• Hyperparameter Optimization: Further tuning could improve model performance. Techniques like grid search and Bayesian optimization can accelerate this process.
• Data Augmentation: Adding more data augmentation techniques could enhance the model’s generalization capabilities and robustness against overfitting.
• Transfer Learning: Applying transfer learning from larger datasets, such as ImageNet, could improve feature extraction and overall performance.
• Adversarial Training: Enhancing model robustness against adversarial attacks by incorporating adversarial cases during training.
• Model Compression: Techniques like pruning and quantization could be explored to deploy the model on edge devices with limited computational resources.
• Broader Dataset Application: Testing the model's adaptability on different image classification tasks by evaluating it on datasets other than CIFAR-10.
• Interpretability and Explainability: Developing methods for interpreting the model’s decisions to foster transparency and trust in critical applications.

Contributing

Contributions are welcome! Feel free to submit a pull request or open an issue to suggest improvements or report bugs.