CIFAR-10 Image Classification with Custom ResNet

Project Overview

The goal is to develop a deep learning model that can effectively classify images from the CIFAR-10 dataset while adhering to specific constraints:

• Model size under 5 million parameters
• Test accuracy target of 80% or higher

Dataset

CIFAR-10 consists of:

• 60,000 32x32 color images
• 10 different classes
• 50,000 training images and 10,000 test images
• Classes: airplane, automobile, bird, cat, deer, dog, frog, horse, ship, truck

Method

Our approach uses a custom ResNet architecture with several key optimizations:

Architecture

• Custom ResNet with progressive channel expansion (3→32→64→128→256)
• Residual blocks with dual convolution layers and skip connections
• Total parameters: 2.76M (well under 5M limit)
• Integrated dropout (0.3) for regularization

Training Strategy

• Advanced data augmentation pipeline including MixUp (α=0.2)
• RAdam optimizer with adaptive learning rate scheduling
• Label smoothing (0.1) and weight decay (1e-4)
• Comprehensive regularization techniques

Results and Conclusion

The model demonstrated strong performance:

• Achieved 90.24% validation accuracy
• Successfully maintained parameter count under 5M (2.76M)

References

1. He, K., Zhang, X., Ren, S., & Sun, J. (2016). Deep residual learning for image recognition. CVPR 2016.
2. CIFAR-10 dataset: https://www.cs.toronto.edu/~kriz/cifar.html

Contributors

• Puneeth Kotha [email protected]
• Chenna Kesava Hemanth Reddy Narala [email protected]
• Sai Sandeep Mamidala [email protected]