1) The document describes Ricardo Ñanculef Alegria's PhD project on incremental support vector learning models and algorithms.
2) The goal of the project is to develop theoretical and practical tools to design incremental machine learning methods that can handle constraints on model size/complexity and changes in patterns over time for support vector algorithms.
3) The main hypothesis is that both issues of handling constraints and changes can be efficiently addressed using a new support vector learning architecture based on multiple prototypes or models handling different parts of the feature space.
"Incremental Support Vector Learning Models and Algorithms ...
1. Departamento de Informática
Investigación y Postgrado EX UMBRA
IN
SOLEM
CANDIDATO A DOCTOR: SR. RICARDO ÑANCULEF ALEGRÍA
Título del Proyecto de Tesis Incremental Support Vector Learning Models and
Algorithms
Nombre del Alumno RICARDO ÑANCULEF ALEGRÍA
Fecha de Ingreso al Programa 2do. Semestre año 2006
Pregrado Ingeniero Civil en Informática, UTFSM, Enero 2006.
(Título o Grado Institución, Año)
Prof. Supervisor de Tesis Dr. Héctor Allende Olivares
Fecha Aprobación Tema de Tesis y 25-03-2009
Examen de Calificación
Fecha tentativa de Término 30-12-2010
Comisión Evaluadora Supervisor de Tesis: Dr. Héctor Allende O.
Co-Guía de Tesis: Dr. Claudio Moraga R. Univ.
Dortmund-Alemania y Centro de
Sof-Computing- Mieres, España.
Correferente DI: Dr. Carlos Castro V.
Correferente Externo: Dr. Gonzalo Acuña,
Depto. Ing. Informática –USACH.
2. Departamento de Informática
Investigación y Postgrado EX UMBRA
IN
SOLEM
ABSTRACT:
Machine learning algorithms are usually designed under the so called batch paradigm, where all
the available examples are provided to the system simultaneously and used as times as desired
in the learning process. Contrasting with this scenario, the purpose in this project is the design
of algorithms under the incremental paradigm, where examples are provided step by step,
continuously on time, and are not arbitrarily available for learning in further steps. Incremental
learning fits better with a natural notion of learning, in which the learner is constantly adapting
to new information from the environment. However, this scenario also means new challenges
with respect to the batch setting since the learner has restricted visibility on the learning task
and the patterns can constantly changing along time. Moreover, real-world applications
requiring incremental models usually impose hard computational conditions, such as real time
response and large volumes of data.
Focusing on a family of methods commonly known as support vector algorithms (SVA's) the
aim of this project is to provide theoretical and practical tools to design incremental learning
methods under two main concerns: constraints on the size/complexity of the model which is
stored and continuously updated by the learner (budgetary constraints), and changes in the
patterns extracted by the learner (drift). Handling budgetary constraints requires the ability of
the learner to quantitatively relate the magnitude of the constraint with the expected
performance of the obtained model.
Addressing drift on the other hand, requires the ability to distinguish real structural change from
noise and react consistently. The main hypothesis of this research is that both issues can be
efficiently addressed using a new architecture for support vector learning based on multiple
prototypes or models handling different parts of the feature space.
Finally, due to the actual state of realistic applications, this project is proposed to cover methods
for diverse machine learning tasks including pattern recognition, regression and label ranking.
The developed models will be experimentally validated using a reasonable collection of real-
world and synthetic data sets, assessing both prediction accuracy and computational efficiency.