Analysing the Performance of a Tomographic Reconstructor with Different Neural Networks Frameworks

Abstract

Correction of atmospheric turbulences with the use of guide stars as reference, is one of the most relevant issues of adaptive optics (AO). This is addressed with tomographic techniques such as Multi-object adaptive optics (MOAO). Next generations of extremely large telescopes, will require improvements in computational capabilities of real time control systems. An improved version of CARMEN, a tomographic reconstructor based on machine learning, is presented here. The performing time of two dedicated neural network frameworks, Torch and Theano, is compared, with significant improvements on the training and execution times of the neural networks due to calculations on GPU. Also, the differences between both frameworks are discussed.