A Cognitive Study of Learning with Labeled and Unlabeled Data
Abstract
In semi-supervised learning (SSL) the learner is presented with both labeled and unlabeled data. If the learner makes certain assumptions regarding the distribution of the unlabeled items p(x) and the class conditional p(y | x), they can learn the concept faster and more accurately. We investigate how humans are affected by unlabeled data in a supervised categorization task. Our project lead to better understanding of human learning, improvements in human teaching strategy, improvements in human/machine cooperative learning and improvements in machine learning models. Our empirical evidence for human SSL includes several human behavioral studies that definitively show the influence of unlabeled data in human category learning. Our theoretical models produce plausible semi-supervised learning models for human learning and machine learning. We utilize these observations to enhance human learning.
Document Details
- Document Type
- Technical Report
- Publication Date
- Jan 01, 2012
- Accession Number
- ADA565197
Entities
People
- Timothy T. Rogers
- Xiaojin Zhu
Organizations
- University of Wisconsin–Madison