Data Selective Deep Neural Networks For Image Classification
Marcele O. K. Mendonça, Jonathas O. Ferreira, Paulo S. R. Diniz · 2021 29th European Signal Processing Conference (EUSIPCO) · 2021
As the volume of data keeps growing, the use of deep neural networks has been widespread in a variety of applications, including image classification. This big available data has led to an increasing interest in designing more efficient systems. Most applications use all training data without taking into account their relevance. A mini-batch gradient descent algorithm is preferable in practice, but the chosen batch size is typically based on empirical tests, and it depends on the dataset characteristics. This work proposes a data-selection strategy applied to classification problems leading to computational savings and, in most cases classification error reduction. A few examples corroborate the effectiveness of the proposed approach.