Analysis of instance selection algorithms on large datasets with Deep Convolutional Neural Networks

Deep Convolutional Neural networks (ConvNets) have achieved impressive results in several applications of computer vision and speech processing. With the availability of a large training set, it is common to find that the set contains useless samples (instances), either redundant or noisy. The process of removing these instances is called instance selection in the machine learning field. This paper evaluates the effectiveness of reducing the number of instances in the training set on deep (ConvNets) using a variety of instance selection methods to speed up the training time. We then study how these methods impact on classification accuracy. Moreover, many instance selection methods require a long running time for obtaining a representative subset of the dataset, especially if the dataset is large with high dimensionality. One of the popular algorithms in instance selection is Random Mutation Hill Climbing (RMHC). We propose a new approach in order to make RMHC work much faster with the same accuracy compared to original RMHC.