Doctoral dissertation
Methodology for adapting machine learning and large-scale optimization problems to be solved by gradient-based algorithms
Views: 0 | Downloads: 0
Abstract
The work presented in this thesis deals with the yet unrealized potential of gradient-based
algorithms in the fields of computer science other than deep learning. Immense advances
in deep learning have revealed the power of both automatic differentiation and gradient
descent for solving large-scale optimization problems in a scalable and computationally
efficient way. However, its use is still extremely limited due to the lack of methodologies to
bring general optimization problems on a similar footing to that of deep learning, as well
as due to a belief that deep learning problems are somehow special and that the methods
employed there do not transfer well to other problems in computer science that include
optimization. In this thesis, we challenge this belief by demonstrating the usefulness of
gradient-based algorithms on a diverse set of real-world problems from machine learning
and large-scale optimization. One of the goals of this thesis is to show the universality and
usefulness of gradient-based algorithms even in the fields where it is not clearly obvious
how to apply them. To this aim, we have set out to focus on cases where the applicability
and efficiency of gradient-based algorithms seem especially surprising, such as discrete
optimization, clustering, and problems that include simulations.
In this thesis, we develop methodological frameworks for applying gradient-based algorithms
in machine learning and large-scale optimization by adapting (transforming) those
problems to a form where generic gradient-based algorithms can be freely applied. These
problem adaptations include developing viable ways to handle constraints, transforming
problems by using a more suitable encoding, and rewriting problems to a form that allows
gradient calculation (even for simulations and discrete problems). With the aim of comprehensibility,
we tackle problems in all three branches of machine learning (supervised,
unsupervised, and reinforcement learning) and both discrete and continuous large-scale
optimization. We conclude the thesis with one of the open topics in gradient-based optimization,
adaptive parameter control of gradient descent. We propose a novel, algorithmagnostic
approach to adapt gradient descent parameters in an online fashion.