Tese de Doutorado em Inteligência Artificial

UNIVERSIDADE FEDERAL DO RIO GRANDE DO SUL

INSTITUTO DE INFORMÁTICA

PROGRAMA DE PÓS-GRADUAÇÃO EM COMPUTAÇÃO

———————————————-

DEFESA DE TESE DE DOUTORADO

Aluno: Sandro Rama Fiorini

Orientador: Profa. Dra. Mara Abel

Título: Similarity, Structure and Spaces: Representation of Part-Whole Relations in Conceptual Spaces

Linha de Pesquisa: Inteligência Artificial

Data: 29/05/2014

Horário: 11:00h

Local: Prédio 43412 – Sala 215 (Sala de videoconferência), Instituto de Informática

Banca Examinadora:

Prof. Dr. Giancarlo Guizzardi (UFES) p/videoconferência

Prof. Dr. Luis da Cunha Lamb (UFRGS)

Prof. Dr. Marcelo Finger (USP)

Presidente da Banca: Profa. Dra. Mara Abel

Resumo:

One of main remaining challenges in Artificial Intelligence is how to make intelligent systems to ground high-level abstractions in reality. At least part of this problem can be drilled down to the question of what representation framework is best suited in a way that facilitates object recognition. Animal cognition, particularly in humans, can give a clue of how such representation framework looks like. According to studies on the cognition of object recognition, humans usually process incoming stimuli by analysing it holistically (whole) and structurally (parts). Also, object recognition involves similarity judgements between incoming stimuli and recorded patterns. This suggests that representation in cognition incorporates, in one or more levels, similarity and holistic-structural information about concepts.

We hold the view that computer systems dealing with part-whole relations should be able to represent similarity and holistic-structural information. However, there exists no representation framework that allows such information to be represented in an integrated way.

This thesis proposes a concept representation framework for representing information about similarity between part-whole structures. We base our proposal on the theory of conceptual space, focusing on prototypical structures in part-whole relations. Traditional symbolic approaches for representing part relation representation do not account for prototypical structures. In our proposal, parts and wholes are represented in distinct conceptual spaces called holistic and structure spaces. Holistic spaces allow for systematic similarity judgements between wholes. On the other hand, structure spaces allow for systematic similarity judgement between structures of parts. A point in the structure space denotes a particular part structure; regions in the space represent different general types of part structures. By redefining conceptual spaces as a product of holistic and structure spaces, our goal is to allow one to represent similarity information between different wholes, taking into consideration the similarity between shared parts and their configurations. We also argue that the structural space can represent prototype effects: structural types are formed around typical arrangements of parts. We also show how structure space captures the variations in part structure of a given concept across different domains. In addition, we discuss how some taxonomies of part-whole relations can be understood within our framework.

We also provide a mathematical formulation of holistic-structure spaces. It is a specialization of Gardenfors’ theory of conceptual spaces and Aisbett and Gibbon’s formulation of the same theory based on pointed metric spaces. In our work, a whole and its parts are represented as regions in metric spaces and their relationship is given by their product. The resulting formulation is intended as a basis for concept representation in computer vision programs.

The result of this thesis is a concept representation framework that allows one to represent similarity relations between conceptual entities, taking into account their partonomical structure. We believe this work can be particularly useful in object recognition systems and other computer applications dealing with low-level data. The benefits should be two-fold: in one side, this framework provides semantic grounding for symbols related to parts in symbol-based computer systems; in the other side, it presents a way of organizing the output of non-symbolic system in a more conceptual structure.

Palavras-chave: part-whole relations; conceptual spaces; similarity; concept representation