Tensor Train Singular Regularisation
The framework allows us to apply various methods for regularising singular numbers with less computational cost
About the framework
Applying TT decomposition to any network
Replacing convolutional layers decomposed with Tensor Train
Calculating singular numbers of convolutional layers
Implemented exact and degree methods, as well as regularisation methods based on them
Robust metrics calculation
Determines the robustness of the network to adversarial attacks
Procedures for training convolutional network architectures on the classification task
Applied to decomposed networks together with regularisation methods
For whom?
Application and software developers will be able to solve the image classification problem more efficiently
Researchers in ML and robustness of neural network training will help speed up computation
Companies will be able to reduce the cost of computing resources
Advantages
Accelerated network regularisation
Increased network stability
Network compression
