Transitional Information Extraction
Author(s)
Benkert, Ryan
Advisor(s)
Editor(s)
Collections
Supplementary to:
Permanent Link
Abstract
The practical deployment of neural networks is contingent on their ability to extract information from input data beyond a predictive output. Unfortunately, neural networks do not naturally implement this property and are explicitly optimized to compress information unrelated to the target application. For this purpose, extracting information from neural networks is frequently coupled with additional application constraints or with utilizing a collection of neural networks (or weights) that require several iterative forward passes. While each of these paradigms show promising results, each comes with a set of drawbacks that limit their deployment in the real world. Specifically, implementing constraints can result in a performance degradation on the application and iterative approaches require several forward passes which is undesirable in settings with limited computational resources.
In this thesis, we address these limitations by realizing iterative methods as a sub-category of a broader family of paradigms, namely transitional information extraction. A key observation we make in this thesis is that the source representations are generalizable beyond several forward passes and can be implemented within a single neural network. Our approach is based on a detailed analysis of existing paradigms where we identify causes of performance decline for application constraints, as well as implication of limited data on the information extraction capabilities of the network. Our study results in a suite of algorithms that implement transitional information extraction in the fields of uncertainty estimation, performance regression, robustness, and underspecified settings.
Sponsor
Date
2024-04-29
Extent
Resource Type
Text
Resource Subtype
Dissertation