Learning Approximate Distribution-Sensitive Data Structures

Zenna Tavares, Armando Solar-Lezama

Abstract:

We present a computational model of mental representations as data-structures which are distribution sensitive, i.e., which exploit non-uniformity in their usage patterns to reduce time or space complexity.
Abstract data types equipped with axiomatic specifications specify classes of concrete data structures with equivalent logical behavior.
We extend this formalism to distribution-sensitive data structures with the concept of a probabilistic axiomatic specification, which is implemented by a concrete data structure only with some probability.
We employ a number of approximations to synthesize several distribution-sensitive data structures from probabilistic specification as deep neural networks, such as a stack, queue, natural number, set, and binary tree.

TL;DR: We model mental representations as abstract distribution-sensitive data types and synthesize concrete implementations using deep networks from specification

Conflicts: mit.edu

Keywords: Unsupervised Learning