Learning concepts through the fusion of representations of distinct visual domains with deep learning

• Grant number: 18/22191-6
• Support Opportunities: Scholarships in Brazil - Scientific Initiation
• Start date: March 01, 2019
• End date: July 14, 2020
• Field of knowledge: Physical Sciences and Mathematics - Computer Science
• Principal Investigator: Moacir Antonelli Ponti
• Grantee: João Guilherme Madeira Araújo
• Host Institution: Instituto de Ciências Matemáticas e de Computação (ICMC). Universidade de São Paulo (USP). São Carlos , SP, Brazil
• Associated scholarship(s): 19/19030-3 - Weight agnostic compositional pattern producing networks and its analysis for learning representations, BE.EP.IC

Abstract

The human cognitive system is able to abstract concepts so that they are related under the same nature. For example, a person alone can be recognized by his or her photo, artistic portrait, caricature, facial composite or even a rough draft. From a computational and pattern recognition point of view, these representations are considered distinct, even though the semantics is the same. The challenge is even greater when there is limited supervision, ie it is a learning problem in a set of unlabeled data or a set with very few labels. The project aims to investigate methods to learn abstract representations from datasets and to relate those representations to other series of distinct domains, but with similar semantics. We will consider feature learning methods, particularly convolutional neural networks, and auto-encoders, and explore architectures capable of achieving a marriage between the learned representations. In addition, each solution will be studied from the point of view of its learning guarantees.