ask_my_thesis / assets /txts /pg_0036.txt
jordyvl's picture
First commit
e0a78f5
raw
history blame
2.23 kB
4
INTRODUCTION
Recognition (ICDAR) [289].
The thesis aims to fill this gap by proposing novel methods for uncertainty
estimation and failure prediction (Part I), and by providing a framework for
benchmarking and evaluating the reliability and robustness of DU technology,
as close as possible to real-world requirements (Part II).
Table 1.1. Comparative analysis of keywords in the ICDAR 2021 proceedings. While
many DU subtasks are represented, there is a lack of keywords related to IA. Do note
that calibration is used in the context of camera calibration, and not in the context of
confidence estimation.
keyword
freq
keyword
freq
document
classification
3388
242
33
0
key information
56
question answering
106
layout analysis
223
calibration/calibrate
temperature scaling
failure prediction
misclassification detection
out-of-distribution
OOD
predictive uncertainty
0
25
0
In the remainder of the Introduction, I will sketch the surrounding research
context, followed by the problem statement and research questions, and finally
the outline of the thesis manuscript.
1.1
Research Context
All chapters of this dissertation have been executed as part of the Baekeland
PhD mandate (HBC.2019.2604) with financial support of VLAIO (Flemish
Innovation & Entrepreneurship) and Contract.fit. The latter is a Belgian-based
software-as-a-service (SaaS) provider of Intelligent Document Processing (IDP)
drawing on innovations in DU to power their product suite (email-routing,
Parble), and my generous employer since 2017.
Some of the joint work (Chapter 5) has been partially funded by a PhD
Scholarship from AGAUR (2023 FI-3-00223), and the Smart Growth Operational
Programme under projects no. POIR.01.01.01-00-1624/20 (Hiper-OCR - an
innovative solution for information extraction from scanned documents) and
POIR.01.01.01-00-0605/19 (Disruptive adoption of Neural Language Modelling
for automation of text-intensive work).
Moreover, given that the dissertation work has been performed over a large
span of time, it warrants putting it in the larger context and dynamics of AI
innovations, the state of DU as a field, how notions of ’reliability’ have evolved
over time, and finally the business context.