Geoscience Reference
In-Depth Information
CHAPTER
7
Human Semi-Supervised
Learning
Suppose a young child is learning the names of animals. Dad occasionally points to an animal and
says “dog!” But most of the time, the child just watches all sorts of animals by herself. Do such passive
experiences help the child learn animals, in addition to the explicit instructions received from Dad?
Intuitively, the answer appears to be “yes.” Perhaps surprisingly, there is little quantitative study on
this question. Clearly, passive experiences are nothing more than unlabeled data, and it seems likely
that humans exploit such information in ways similar to how semi-supervised learning algorithms
in machines do. In this chapter, we demonstrate the potential value of semi-supervised learning on
cognitive science.
7.1 FROMMACHINE LEARNINGTOCOGNITIVE SCIENCE
Humans are complex learning systems. Cognitive Science, an interdisciplinary science that embraces
psychology, artificial intelligence, neuroscience, philosophy, etc., develops theories about human in-
telligence. Traditionally, cognitive science has benefited from computational models in machine
learning, such as reinforcement learning, connectionist models, and non-parametric Bayesian mod-
eling, to name a few. To help understand experiments described in this chapter, we start by providing
a “translation” of relevant terms from machine learning to cognitive science:
Instance
x
: a stimulus, i.e., an input item to a human subject. For example, a visual stimulus
can be a complex shape representing a microscopic pollen particle.
Class
y
: a concept category for humans to learn. For example, we may invent two fictitious
flowers
Belianthus
and
Nortulaca
that subjects are asked to learn to recognize.
Classification: a concept learning task for humans. Given a pollen particle as the visual stimulus,
the human decides which flower it comes from.
Labeled data: supervised experience (e.g., explicit instructions) from a teacher. Given a pollen
particle
x
, the teacher says “this is from Belianthus.”
Unlabeled data: passive experiences for humans. The human subject observes a pollen particle
x
without receiving its true class.
Learning algorithm: some mechanism in the mind of the human subject. We cannot directly
observe how learning is done.
