ARTIFICIAL NEURAL NETWORKS COMPRESSION SYSTEM BASED ON ITERATIVE APPLICATION OF TENSOR APPROXIMATIONS Russian patent published in 2020 - IPC G06N3/02 

FIELD: physics.

SUBSTANCE: invention relates to compression of artificial neural networks. System comprises a compression device comprising a module for automatic determination of compression parameters (rank selector module) and a module which performs replacement of parameters of convolutional/fully connected layers of the NN with their low-rank approximation, obtained using tensor/matrix expansions (tensor approximator module), and a fine tuning device, wherein the compression device receives to the input of the NN, the rank selector module automatically for each convolutional/fully connected layer of the NN selects the rank of the tensor decomposition, which is used when approximating the weight tensor, after which the tensor approximator module changes the weight of the layer to its low-rank approximation such that the total number of parameters of the new tensors is less than the number of parameters in the initial tensor, and fine adjustment device receives input of converted NN from compression device and outputs to output optimized NN, having better predictive ability due to correction of model parameters, which is performed by method of back propagation of error using database.

EFFECT: technical result consists in improvement of compression efficiency of artificial neural networks.

1 cl, 5 dwg