RM Roles

Can we encode tree structures into Tensor Product Representations (TPR’s) where the rank of the TPR doesn’t grow with the depth of the tree?

Can we encode tree structures into Tensor Product Representations (TPR’s) where the rank of the TPR doesn’t grow with the depth of the tree?  This project develops a new technique called Recursive Matrix Roles to generate child roles from a given parent role (and with the same dimensionality as the parent role).  After allocating a fixed set of roles for the first level children, this technique allows the remaining child roles of a tree to be dynamically allocated in a predictable (and learnable) way.  Two proof-of-concept examples are developed using the technique: beta reduction and tree adjoining.