Multi-scale tensor network architecture for machine learning

Abstract We present an algorithm for supervised learning using tensor networks, employing a step of data pre-processing by coarse-graining through sequence wavelet transformations. These transformations are represented as set network layers identical to those in multi-scale entanglement renormalization ansatz network. perform and regression tasks model based on matrix product states (MPSs) acting the coarse-grained data. Because entire consists contractions (apart from initial non-linear feature map), we can adaptively fine-grain optimized MPS ‘backwards’ with essentially no loss performance. The itself is trained adaptive density group algorithm. test our methods performing classification task audio temperature time-series data, studying dependence training accuracy number showing how fine-graining may be used initialize models which access finer-scale features.