• A
  • A
  • A
  • ABC
  • ABC
  • ABC
  • А
  • А
  • А
  • А
  • А
Regular version of the site

Book chapter

Designing a Neural Network Primitive for Conditional Structural Transformations

Demidovskij A., Babkin E.

Among the problems of neural network design the challenge of explicit representing conditional structural manipulations on a sub-symbolic level plays a critical role. In response to that challenge the article proposes a computationally adequate method for design of a neural network capable of performing an important group of symbolic operations on a sub-symbolic level without initial learning: extraction of elements of a given structure, conditional branching and construction of a new structure. The neural network primitive infers on distributed representations of symbolic structures and represents a proof of concept for the viability of implementation of symbolic rules in a neural pipeline for various tasks like language analysis or aggregation of linguistic assessments during the decision making process. The proposed method was practically implemented and evaluated within the Keras framework. The network designed was tested for a particular case of transforming active-passive sentences represented in parsed grammatical structures.

In book

Designing a Neural Network Primitive for Conditional Structural Transformations
Edited by: S. Kuznetsov, A. I. Panov, K. Yakovlev. Switzerland: Springer, 2020.