Receptive field resolution analysis in convolutional feature extraction

Abstract

Instead of introducing new learning algorithm for solving complex classification tasks, many research groups in machine learning have focused on creating a good feature representation. In addition, labeled data is often difficult and expensive to obtain sufficiently large amount of data. So, learning features from unlabeled data is proposed since unlabeled data is much easier to obtain than the labeled data. In this work, a highlevel feature representation is created by a sparse autoencoder with convolutional extraction. A sparse autoencoder is an unsupervised feedforward neural network that is trained to predict the input itself and has been widely used for learning good feature representation. A major advantage of this feature extraction approach not only provides good feature representation for higherlevel tasks but also can scale up to large images. However, there are several parameters that require careful selection to obtain high performance. The main objective in this work is to present a detailed analysis on the effect of receptive field resolution in handwritten classification based on MNIST database. In the experiment, the results show that receptive field resolution is one of the critical parameters to achieve state-of-the-art performance.