Skip to content

A trained CNN model that is feature extracted and fine-tuned from a pre-trained VGG16 model on the image-net dataset using SVD-augmented images

Notifications You must be signed in to change notification settings

kankenny/SVDNet

Repository files navigation

On Image Data Augmentation: Training CNNs with Low-Rank Approximation of Images using the Singular Value Decomposition

Kennette Basco
New York Institute of Technology
Department of Engineering and Computer Science

SVDNet

SVDNet is a feature extracted and fine-tuned pre-trained VGG16 model using SVD-augmented images

MNIST PATCH MALARIA CATSDOGS

ABSTRACT

This study investigates enhancing the generalization of Convolutional Neural Networks (CNNs) for low-resolution images using Singular Value Decomposition (SVD) augmented images. It provides a theoretical overview of SVD and a method to control compression through hyperparameters: energy factor and skip threshold. Methods involve preprocessing, training models, and benchmarking validation and testing accuracy on various datasets (grayscale and RGB). Experiments using SVD augmentation show degraded validation and testing accuracy. In summary, while SVD augmented data shows theoretical promise, it does not alleviate the training-production data skew and even increases the signal-noise ratio. Rigorous hyperparameter tuning of the energy factors and skip threshold may regularize CNNs, but the computational resources required outweigh the marginal gains in regularization, if at all. This study highlights the limitations of SVD based augmentation and underscores the need for alternative regularization techniques in deep learning. Keywords: convolutional neural networks, singular value decomposition, regularization, adversarial learning.

Other Resources

About

A trained CNN model that is feature extracted and fine-tuned from a pre-trained VGG16 model on the image-net dataset using SVD-augmented images

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published