Multi-objective optimization in multi-task learning

• Grant number: 14/13533-0
• Support type: Scholarships in Brazil - Doctorate
• Effective date (Start): September 01, 2014
• Effective date (End): August 31, 2018
• Field of knowledge: Physical Sciences and Mathematics - Computer Science
• Cooperation agreement: Coordination of Improvement of Higher Education Personnel (CAPES)
• Principal researcher: Fernando José von Zuben
• Grantee: Marcos Medeiros Raimundo
• Home Institution: Faculdade de Engenharia Elétrica e de Computação (FEEC). Universidade Estadual de Campinas (UNICAMP). Campinas , SP, Brazil

Abstract

This doctoral research project proposes two new solution perspectives to multi-task learning. In the literature, multiplicative and additive factors are generally defined a priori to establish the existing inter-relations among mono-task learning processes, so that this joint treatment promotes gain in performance. In this project proposal, the decision of how to jointly explore multiple mono-task learning processes is taken only after obtaining a set of efficient solutions, which represents alternative trade-offs (in other words, alternative configurations of the additive or multiplicative factors mentioned above) among mono-task learning processes. Thus, the decision maker is able to choose the best trade-off after getting more familiar with the existing inter-relations of the mono-task learning processes, or eventually he/she may choose to build a committee machine, i.e. an ensemble of efficient solutions. This multi-objective approach for multi-task learning is novel and is composed of three phases: (1) Formalization of the multiple objectives, each one associated with one single process of mono-task learning; (2) Use of multi-objective optimization in order to obtain a diversified set of efficient solutions, each one representing a distinct trade-off among multiple mono-task learning processes; (3.1) Selection of the best trade-off, according to the preference of the decision maker; or (3.2) Use of an ensemble approach, aggregating multiple trade-off solutions. The proposal of this thesis can be implemented to deal with a diverse set of machine learning problems, including regression and classification. The case studies to be considered should disclose the benefits of the multi-objective approach for multi-task learning, as well as promote performance comparison with other multi-task learning approaches in the literature. (AU)

Scientific publications

(References retrieved automatically from Web of Science and SciELO through information on FAPESP grants and their corresponding numbers as mentioned in the publications by the authors)

RAIMUNDO, MARCOS M.; DRUMOND, THALITA F.; MARQUES, ALAN CAIO R.; LYRA, CHRISTIANO; ROCHA, ANDERSON; VON ZUBEN, FERNANDO J. Exploring multiobjective training in multiclass classification. Neurocomputing, v. 435, p. 307-320, MAY 7 2021. Web of Science Citations: 0.

RAIMUNDO, MARCOS M.; FERREIRA, V, PAULO A.; VON ZUBEN, FERNANDO J. An extension of the non-inferior set estimation algorithm for many objectives. European Journal of Operational Research, v. 284, n. 1, p. 53-66, JUL 1 2020. Web of Science Citations: 0.

MARQUES, ALAN CAIO R.; RAIMUNDO, MARCOS M.; CAVALHEIRO, ELLEN MARIANNE B.; SALLES, LUIS F. P.; LYRA, CHRISTIANO; VON ZUBEN, FERNANDO J. Ant genera identification using an ensemble of convolutional neural networks. PLoS One, v. 13, n. 1 JAN 31 2018. Web of Science Citations: 3.

Academic Publications

(References retrieved automatically from State of São Paulo Research Institutions)

RAIMUNDO, Marcos Medeiros. . 2018. Doctoral Thesis - Universidade Estadual de Campinas, Faculdade de Engenharia El?trica e de Computa??o.