The Importance of Data Quality in Training a Deep Convolutional Neural Network

It is an acknowledged fact that good quality and varied training data samples are essential for achieving high rates of success in the operation of a convolutional neural network. The performance of the trained network is highly dependent on the training data being as diverse and representative as possible and each learnable feature being present in a large number of training data samples. There are situations when it is not possible to accrue a sufficient number of original samples and artificial sample generation is needed. The current work studies the impact on performance of a convolutional neural network of original training data supplemented with artificially generated samples. Multiple batches of training data consisting of varied combinations of fractions between original and artificial samples have been used to train a convolutional neural network in order to determine the percentage of original data needed to achieve minimal loss of performance.