Use this URL to cite or link to this record in EThOS: https://ethos.bl.uk/OrderDetails.do?uin=uk.bl.ethos.756344
Title: Experimental analysis of representation learning systems
Author: Sanchez Carmona, Vicente Ivan
ISNI:       0000 0004 7429 2974
Awarding Body: UCL (University College London)
Current Institution: University College London (University of London)
Date of Award: 2018
Availability of Full Text:
Access from EThOS:
Full text unavailable from EThOS. Please try the link below.
Access from Institution:
Abstract:
Studying a subject is central to understanding its behavior and what it has learned. In this thesis, we study specific aspects of five representation learning systems for natural language processing tasks. Representation learning systems are a type of machine learning system dedicated to learn representations of data suitable for other machine learning systems, such as classifiers, to operate upon them. Thus, understanding the behavior of and the abilities learned by representation learning systems is crucial for improving the results on the tasks they are used. The aspects on which we focus are interpretability, robustness, and abilities learned. We are interested in obtaining explanations that allow us to understand how a system makes a decision, what factors from the data and internal to the system affect its robustness, and to what extent it has learned a linguistic ability. To do so, we propose to carry out three types of analyses, namely functional, behavioral, and internal analyses which we link with work on the cognitive science, behavioral science, and neuroscience. We present three case studies. In the first study, we provide a functional explanation of a matrix factorization system that allow us to understand how this system makes a prediction. In our second study, we investigate how robust are three systems when the input data suffers a simple transformation and how certain external and internal factors influence their behavior; these systems are trained for the task of natural language inference. Finally, our third study shows that we are able to extract hypernymy from the word embeddings of a popular ReLe system, while studying the influence that the choice of hypernymy dataset plays in the task. In summary, we advance towards better understanding ReLe systems by providing explanations of their predictive behavior and investigating abilities learned by these systems.
Supervisor: Not available Sponsor: Not available
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral
EThOS ID: uk.bl.ethos.756344  DOI: Not available
Share: