Ph.D Thesis

Ph.D StudentRichman Oran
SubjectUncertainty Management in Learning Problems
DepartmentDepartment of Electrical and Computer Engineering
Supervisor PROF. Shie Mannor


Machine learning techniques are used to transform large amount of data into

decisions. This data is often imperfect, either since the sensors are noisy or

since some data is missing or corrupted. Many factors impact the quality

of the data gathered, some of them are under our control. Often lack of

resources is one of the main factors impacting data imperfections, This can

be money (limiting the quality of sensors), bandwidth, power, manpower (for

manual correction of data) or any other resource. This research is dealing

with efficient resource allocation in scenarios where data is used for classification

tasks. Resources can be allocated between different sensors or between

the classified data samples. We focus on binary classification problems due to their vast use in machine learning literature. This research present both practical algorithms

for efficient use of resources as well as some theoretical guarantees. Simulation results with real life data is presented in order to demonstrate the algorithms potential benefit. It can be seen that approximately 30% increase in performance can be achieved using the same resources. We further show that the same concepts can be used to tackle other challenges in

machine learning, namely meta-learning of better classifiers (by reallocation

of false-alarms over the data-space) and on-line feature selection using a novel

multi-arm bandit (MAB) based algorithm.