Tensorization of neural networks for improved privacy and interpretability

José Ramón Pareja Monturiol presents a method that converts black-box functions into tensor-train representations using sketching, cross interpolation and a small sample of the relevant domain. For neural networks, this enables privacy-oriented obfuscation and interpretable estimates of topological phases. The method also supplies an efficient initialisation for tensor-train optimisation and improves the memory–computation trade-off in model compression.