assets/txts/pg_0043.txt

Chapter 2

Fundamentals
This chapter provides all the necessary background knowledge necessary to
understand the contributions of this thesis.
The key questions covered here are:
i.
ii.
iii.
iv.
v.
vi.

How to feed a document to an algorithm to perform arbitrary tasks on it?
How to model language, vision, layout or structure?
How does it learn and then operate at inference time?
How does it estimate prediction uncertainty?
How to evaluate its performance?
How to integrate it as a useful, end-to-end system in a document workflow?

Section 2.1 explains the basic setting from the perspective of statistical learning
theory [472], which is a mathematical framework for analyzing how algorithms
learn from data with minimal error. Section 2.2 gives a primer on reliability and
robustness, particularly calibration, failure detection and relevant evaluation
metrics. Section 2.3 surveys the DU field, and discusses the state of the art in
DU technology. Finally, Section 2.4 covers Intelligent Automation to illustrate
how solving the challenges posed in this thesis will enable to augment human
intelligence, creativity and productivity in straight-through business processes.

11