Comparative Study of Neural Network Frameworks for the Next Generation of Adaptive Optics Systems

Abstract

Many of the next generation of adaptive optics systems on large and extremely large telescopes require tomographic techniques in order to correct for atmospheric turbulence over a large field of view. Multi-object adaptive optics is one such technique. In this paper, different implementations of a tomographic reconstructor based on a machine learning architecture named “CARMEN” are presented. Basic concepts of adaptive optics are introduced first, with a short explanation of three different control systems used on real telescopes and the sensors utilised. The operation of the reconstructor, along with the three neural network frameworks used, and the developed CUDA code are detailed. Changes to the size of the reconstructor influence the training and execution time of the neural network. The native CUDA code turns out to be the best choice for all the systems, although some of the other frameworks offer good performance under certain circumstances.

Multi-GPU Development of a Neural Networks Based Reconstructor for Adaptive Optics

“…For RAM execution has almost a linear relation, where doubling or tripling the number of GPUs has the same impact in times. Also, as it was analyzed in previous works [24,25], loading data from SSD instead of directly from RAM has an important impact on performance, although in this case it could be observed how this difference is almost fixed (about 0.3 milliseconds) for every number of GPUs.…”

“…For these analysis, only three of the most popular [42][43][44] have been selected. Although it was used in our previous papers [24,26], Theano has been left out of the comparative due to the recent announcement of their creators not to continue its development [45].…”

“…The training benchmark is the same method used in [24][25][26] and will be summarized. In this case, not only learning rate, momentum, and backpropagation algorithm, Stochastic Gradient Descent (SGD) [48], are fixed, but also the minibatch size will always be 256.…”

“…ANN and deep learning have become very popular in recent years [22], and several frameworks have been developed to help researchers in their projects [23]. Most of these frameworks provide GPU acceleration, and some of them have shown good results speeding up CARMEN training and execution [24][25][26]. However, only one GPU has been used in previous tests, while there are already some Multi-GPU implementations both for convolutional neural networks [27,28] and for L&A [29].…”

An approach using deep learning for tomographic reconstruction in solar observation

“…The use of AO, implies the necessity of a reconstruction system to compensate the aberrations introduced by the atmosphere [6]. Machine learning techniques as SOM [7] or MARS [8] have been used in nocturnal observations with success [9], but the use of neural networks [10] has been proved to be a better solution [11].…”