1-D CNNs with lock-free asynchronous adaptive stochastic gradient descent algorithm for classification of astronomical spectra

At present, large-scale sky surveys have obtained a large volume of stellar spectra. An efficient classification algorithm is of great importance to the practice of astronomical research. In this paper, we propose a novel parallel optimization algorithm based on a lock-free and shared-memory environment to solve the model for astronomical spectra class. Firstly, the SMOTE-TOMEK and RobustScaler are introduced to use for class balancing and data normalization. Secondly, 1-Dimensional Convolutional Neural Networks (1-D CNN) with L2-norm loss function is utilized as a classifier. Finally, LFA-SGD, LFA-Adagrad, LFA-RMSprop and LFA-Adam algorithms are proposed and applied to the classifier solution. The Lock-Free and shared-memory parallel Asynchronous environment (LFA) relies on GPU multiprocessing, allowing the algorithm to fully utilize the multi-core resources of the computer. Due to its sparsity, the convergence speed is significantly faster. The experimental results show that LFA-SGD algorithm and its variants achieved state-of-the-art accuracy and efficiency for astronomical spectra class.