assets/txts/pg_0044.txt

12

FUNDAMENTALS

Contents
2.1

2.2

2.3

2.4

2.1

Statistical Learning - basics . . . . . . . . . . . .
2.1.1 Neural Networks . . . . . . . . . . . . .
2.1.2 Probabilistic Evaluation . . . . . . . . .
2.1.3 Architectures . . . . . . . . . . . . . . .
Reliability and Robustness . . . . . . . . . . . .
2.2.1 Generalization and Adaptation . . . . .
2.2.2 Confidence Estimation . . . . . . . . . .
2.2.3 Evaluation Metrics . . . . . . . . . . . .
2.2.4 Calibration . . . . . . . . . . . . . . . .
2.2.5 Predictive Uncertainty Quantification . .
2.2.6 Failure Prediction . . . . . . . . . . . . .
Document Understanding . . . . . . . . . . . . .
2.3.1 Task Definitions . . . . . . . . . . . . . .
2.3.2 Datasets . . . . . . . . . . . . . . . . . .
2.3.3 Models . . . . . . . . . . . . . . . . . . .
2.3.4 Challenges in Document Understanding
Intelligent Automation . . . . . . . . . . . . . .

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

12
14
15
17
18
19
20
21
25
27
29
30
31
33
34
35
38

Statistical Learning

Two popular definitions of Machine Learning (ML) are given below.
Machine Learning is the field of study that gives computers the ability
to learn without being explicitly programmed. [406]
A computer program is said to learn from experience E with respect to
some class of tasks T, and performance measure P, if its performance
at tasks in T, as measured by P, improves with experience E. [317]
Following these, different types of learning problems [472] can be discerned, of
which the most common (and the one used throughout our works) is supervised
learning. It defines experience E as a set of input-output pairs for which the
task T is to learn a mapping f from inputs X ∈ X to outputs Y ∈ Y, and the
performance measure P is the risk or expected loss (Equation (2.1)), given a
(0-1) loss function ` : Y × Y → R+ .
R(f ) = E(X,Y )∼P [`(Y, f (X))]

(2.1)

The mapping f (·; θ) : X → Y is typically parameterized by a set of parameters
θ (omitted whenever it is fixed) and a hypothesis class F, which is a set of