2022 · Deep Learning
Convolutional Neural Networks for Image Classification
Deep Learning CNN Computer Vision
Convolutional Neural Networks for Image Classification
BITS Pilani | Sep 2022 – Dec 2022
Python PyTorch CNN CIFAR-10 Computer Vision
Project Overview
Deep dive into convolutional neural network architectures, implementing various designs from scratch and analyzing their performance on standard image classification benchmarks.
Key Contributions
Architecture Implementation: Implemented LeNet, VGG-style, and custom CNN architectures from scratch in PyTorch, achieving 92% accuracy on CIFAR-10
Optimization Experiments: Conducted ablation studies on optimization techniques (SGD, Adam, RMSprop), learning rate schedules, and batch normalization effects
Regularization Analysis: Analyzed impact of dropout, L2 regularization, and data augmentation on model generalization, reducing overfitting gap by 15%
Technologies Used
- Languages: Python
- Framework: PyTorch
- Dataset: CIFAR-10 (60,000 images, 10 classes)
- Hardware: GPU acceleration (CUDA)
Key Results
| Metric | Value |
|---|---|
| Final Accuracy | 92% |
| Dataset | CIFAR-10 |
| Architectures Tested | 4 |
| Overfitting Reduction | 15% |
Architectures Explored
- LeNet-5: Classic architecture, baseline performance
- VGG-style: Deeper networks with 3x3 convolutions
- Custom CNN: Optimized architecture for CIFAR-10
- ResNet-inspired: Skip connections for deeper training