Approximation and learning with compositional tensor trains
arxiv.org·4d
🧠Machine Learning
Preview
Report Post

Title:Approximation and learning with compositional tensor trains

View PDF HTML (experimental)

Abstract:We introduce compositional tensor trains (CTTs) for the approximation of multivariate functions, a class of models obtained by composing low-rank functions in the tensor-train format. This format can encode standard approximation tools, such as (sparse) polynomials, deep neural networks (DNNs) with fixed width, or tensor networks with arbitrary permutation of the inputs, or more general affine coordinate transformations, with similar complexities. This format can be viewed as a DNN with width exponential in the input dimension and structured weights matrices. Compared to DNNs, this format enables…

Similar Posts

Loading similar posts...