Algorithm for Multistage Compression of Convolutional Layers

Algorithms based on neural network models are effective for solving a wide range of problems, including compression of neural network models, which allows efficient use of models with limited resources. Their application allows reducing the total volume of the model and, as a consequence, reducing the computational load. The presented iterative low-rank tensor approximation algorithm outperforms single-stage methods while maintaining the required model accuracy, and its purpose is the compression of any neural network model architecture, which utilizes convolutional and fully connected layers. Testing of the presented algorithm has shown an accuracy advantage of 22-74% compared to the existing ones.