Mental Representations as Distribution-Sensitive Data Structures.

Zenna Tavares, Armando Solar-Lezama

Abstract: We model representations as data-structures which are distribution sensitive, i.e., which exploit regularities in their usage patterns to reduce time or space complexity. We introduce probabilistic axiomatic specifications to extend abstract data structures - which specify a class of representations with equivalent logical behavior - to a distribution-sensitive data structures. We reformulate synthesis of distribution-sensitive data structures as a continuous function approximation problem, using deep neural networks to induce stacks, queues, and natural number representations.