In past decades the topic of self-learning algorithms, such as deep
neural networks, has gained popularity. The idea appeared in the 50s of
the past decade, but hardware and software was an obstacle to research.
Appearance of parallel computing systems and frameworks such as
TensorFlow, Theano and Caffe, providing parallelization at the libraries
level, allowed to accelerate time of learning and inference in dozens of
times when using modern graphic accelerator and CUDA technology.
Four applications were developed: for classification of compressed
images, recognition of sorts of durum wheat, prediction of the level of
environment pollution and two-step approach for the particle track
recognition in GEM detectors. Programs were written using Python,
TensorFlow and Keras libraries.
Performing computations on HybriLIT virtual machine with NVIDIA Tesla
M60 graphic accelerator shows a significant gain of the speed of
learning and inference.