Artificial neural network (ANN) is widely applied as data-driven modeling tool in hydroinformatics due to its broad applicability of handing implicit and nonlinear relationships between the input and output data. To obtain a reliable ANN model, training ANN using the data is essential, but the training is usually taking many hours for large data set and/or for large systems with many variants. This may not be a concern when ANN is trained for offline applications, but it is of great importance when ANN is trained or retrained for real-time and near real-time applications, which are becoming an increasingly interested research theme while the hydroinformatics tools will be an integral part of smart city operation system. Based on author’s previous research projects, which proved that GPU-based ANN is more than 10X efficient than CPU-based ANN for constructing the meta-model (fast simulation), applied as a surrogate of the physics-based model (slow simulation). This paper presents the latest development of GPU-based ANN computing kernels that is implemented with OpenCL an Open Compute Language. The generalized ANN can be used an efficient machine learning library for data-driven modeling. The performance of the implemented library has been tested with the benchmark example and compared with the previous results.
Wu, Zheng Yi, "Portable GPU-Based Artificial Neural Networks For Data-Driven Modeling" (2014). CUNY Academic Works.