Neural networks for adaptive processing of structured data

Abstract

Structured domains are characterized by complex patterns which are usually represented as lists, trees, and graphs of variable sizes and complexity. The ability to recognize and classify these patterns is fundamental for several applications that use, generate or manipulate structures. In this paper I review some of the concepts underpinning Recursive Neural Networks, i.e. neural network models able to deal with data represented as directed acyclic graphs.