Human Semi-Supervised Learning - Introduction to Semi-Supervised Learning - page 78

BIBLIOGRAPHICAL NOTES
Advances in machine learning may shed light on the cognitive process behind human learning,
which may in turn lead to novel machine learning approaches. Such synergy has long been rec-
ognized [ 130 , 105 ]. The study of human semi-supervised learning is still in its infancy. In the
cognitive psychology literature, there have been observations on the effect of unlabeled data to su-
pervised learning, although it was not discussed in a formal semi-supervised learning framework.
For examples, see [ 127 , 137 , 175 , 194 ]. The earliest quantitative study of human learning with
reference to modern semi-supervised machine learning models is [ 166 ]. It used drawings of artificial
fish to show that human categorization behavior can be influenced by the presence of unlabeled
instances. Though certainly suggestive, the experiment had two limitations. First, it used a single
positive labeled example and no negative labeled examples, making it a one-class setting similar
to novelty detection or quantile estimation, instead of classification. Second, since the stimuli are
representations of a familiar concept (i.e., fish), it is difficult to know whether the results reflect prior
knowledge about the category, or new learning obtained over the course of the experiment. The first
clear demonstration of human semi-supervised learning is [ 214 ].
The combination of active learning and semi-supervised learning has attracted attention in
machine learning, see, e.g., [ 85 , 126 , 132 , 183 , 203 , 202 , 213 ]. In cognitive science, quantitative
research on human active learning has just started [ 31 , 103 ]. We expect the study of the combination
to be a fruitful direction.