Dimensionality Reduction Using Variational Autoencoders

Abstract

In the past few years, the problem of processing of big data in less time and with great accuracy has been a major concern in the world of machine learning. Problems like overfitting and underfitting have also been on the troubling side of projects in the years leading to the current stages of artificial intelligence. This model works on the principle of dimensionality reduction, a potential solution to the above problems, with the growth of artificial intelligence and machine learning, data has become one of the most important assets and yet the biggest challenge. Over the years fast and accurate processing of data has been a constant pursuit, and this is where dimensionality reduction comes into play. In dimensionality reduction, authors reduce the dimension of the data keeping the essence and all the important points of the data intact, this leads to faster and equally accurate result. In my project, I have shown the concept of dimensionality reduction on images using the fonts dataset. Variational autoencoders are used to reduce dimension of the images from larger dimensions to smaller dimensions. The model was then checked for the training loss and the validation loss. Also, GPU based approach is used for the upscaling of the model and obtaining competent and satisfying results.