Training Classifiers With a Multi-Grid DMRG Algorithm

We introduce a novel machine learning architecture for the classification of large vector data. The architecture mimics the MERA architecture, with each layer providing a new renormalization "scale" to perform a DMRG-like optimization for the training of the network. We observe a dependence of the accuracy and generalization on the number of layers within the architecture, testing on audio classification datasets. We also modify the algorithm for the prediction of future data points in a time-dependent data set, characterizing its performance by the average absolute error in its prediction.