A Tensor Formalism for Computer Science

Over recent years, tensors have emerged as the preferred data structure for model representation and computation in machine learning. However, current tensor models suffer from a lack of a formal basis, where the tensors are treated as arbitrary multidimensional data processed by a large and ever-growing collection of functions added ad hoc. In this way, tensor frameworks degenerate to programming languages with a curiously cumbersome data model. This paper argues that a more formal basis for tensors and their computation brings important benefits. The proposed formalism is based on 1) a strong type system for tensors with named dimensions, 2) a common model of both dense and sparse tensors, and 3) a small, closed set of tensor functions, providing a general mathematical language in which higher level functions can be expressed. These features work together to provide ease of use resulting from static type verification with meaningful dimension names, improved interoperability resulting from defining a closed set of just six foundational tensor functions, and better support for performance optimizations resulting from having just a small set of core functions needing low-level optimizations, and higher-level operations being able to work on arbitrary chunks of these functions, as well as from better mathematical properties from using named tensor dimensions without inherent order. The proposed model is implemented as the model inference engine in the Vespa big data serving engine, where it runs various models expressed in this language directly, as well as models expressed in TensorFlow or Onnx formats.