I think the reasoning behind the nameing convention is that tensors in DL are just an abstraction of multi-dimensional arrays and that the operations you do on

I think the reasoning behind the nameing convention is that tensors in DL are just an abstraction of multi-dimensional arrays and that the operations you do on multiple arrays such as sum or multiply are algebraic tensor operations due to their multi linearity. It could also be dated to earlier work on computer vision where tensors were used to store a bunch of spatial coordinates so it might be that back then it was even more aligned with their mathematical definition. That said the grief doesn’t seem to be with the naming convention but rather with how tensor classes are designed in common DL frameworks. http://nlp.seas.harvard.edu/NamedTensor http://nlp.seas.harvard.edu/NamedTensor