Geoscience Reference
In-Depth Information
BIBLIOGRAPHICAL NOTES
Semi-supervised learning is a maturing field with extensive literature. It is impossible to cover all
aspects of semi-supervised learning in this introductory topic. We try to select a small sample of
widely used semi-supervised learning approaches to present in the next few chapters, but have to
omit many others due to space. We provide a glimpse to these other approaches in Chapter 8.
Semi-supervised learning is one way to address the scarcity of labeled data. We encourage
readers to explore alternative ways to obtain labels. For example, there are ways to motivate human
annotators to produce more labels via computer games [
177
], the sense of contribution to citizen
science [
165
], or monetary rewards [
3
].
Multiple researchers have informally noted that semi-supervised learning does not
always
help.
Little is written about it, except a few papers like [
48
,
64
]. This is presumably due to “publication
bias,” that negative results tend not to be published. A deeper understanding of when semi-supervised
learning works merits further study.
Yarowsky's word sense disambiguation algorithm [
191
] is a well-known early example of self
training. There are theoretical analyses of self-training for specific learning algorithms [
50
,
80
].
However, in general self-training might be difficult to analyze. Example applications of self-training
can be found in [
121
,
144
,
145
].
