Weight agnostic compositional pattern producing networks and its analysis for learning representations

• Grant number: 19/19030-3
• Support Opportunities: Scholarships abroad - Research Internship - Scientific Initiation
• Start date: December 01, 2019
• End date: February 29, 2020
• Field of knowledge: Physical Sciences and Mathematics - Computer Science - Computing Methodologies and Techniques
• Principal Investigator: Moacir Antonelli Ponti
• Grantee: João Guilherme Madeira Araújo
• Supervisor: Claus de Castro Aranha
• Host Institution: Instituto de Ciências Matemáticas e de Computação (ICMC). Universidade de São Paulo (USP). São Carlos , SP, Brazil
• Institution abroad: University of Tsukuba, Japan
• Associated to the scholarship: 18/22191-6 - Learning concepts through the fusion of representations of distinct visual domains with deep learning, BP.IC

Abstract

Neural networks are relevant tools for learning representations for different tasks. In this context, compositional pattern production networks (CPPNs) are abstractions similar to artificial neural networks that can be trained using similar methods to encode complex life-like patterns. The method canonically used to evolve such networks is neural evolution of augmenting topologies (NEAT), but recently a new method based on reservoir computing, known as weight agnostic neural networks search (WANNS) was developed. This project aims to study the application of WANNS to the evolution of the topology of CPPNs by analyzing the characteristics of the generated patterns, in particular images and soft robots. We believe this can impact on the design of generative networks and the quality of representations. (AU)