Nowadays, the sizes of databases in real-world applications are around TeraByte or PetaByte. Therefore, training neural networks in reasonable times is challenging and requires high-cost computational architectures. OS-ELM is a variant of ELM, proposed for real-world applications.
This algorithm allows training with new data using the previous results without reusing the previous dataset. In this work, we present a parallel model of OS-ELM for classification problems using large-scale databases. The model consists of training several OS-ELM using multithreaded programming. The training dataset is distributed according to the number of working threads. Then, the test dataset is classified by all pre-trained OS-ELMs. Finally, the test dataset is classified using a frequency criterion. Preliminary results show that increasing the number of threads decreases the training time without significantly affecting the test accuracy of each OS-ELM.