Regression Task on Big Data with Convolutional Neural Network

Abstract

As one of most widely utilized methods in deep learning, convolutional neural network (CNN) has been proven effective in many machine learning applications, especially in the areas of image understanding and computer vision. However, CNN is mainly used for applications with the approach of classification, while its usage for regression is not well-studied. In this work, we propose a strategy based on CNN with Visual Geometry Group Network (VGG) for image regression task. We have applied this method on images of MNIST processed with labels of continuous number. In our study, the original discrete classes of handwriting numbers are converted into float numbers with respect to normal distribution, thereby the traditional classification task in MNIST becomes a regression one. In our study, different loss functions such as Mean Absolute Error (MAE) and Log-cosh have been applied and validated. Final results generated by model trained with CNN with VGG with 10-fold cross-validation can be obtained, where MAE is less than 0.25, compared to the much higher error of around 3 with the use of other loss functions and convolutional layers. The significantly reduced error suggests the applicability of our proposed method.