Cognitive Models of Test-Item Effects in Human Category Learning

Imagine two identical people receive exactly the same training on how to classify certain objects. Perhaps surprisingly, we show that one can then manipulate them into classifying some test items in opposite ways, simply depending on what other test items they are asked to classify (without label feedback). We call this the Test-Item Effect, which can be induced by the order or the distribution of test items. We formulate the Test-Item Effect as online semi-supervised learning, and extend three standard human category learning models to explain it.

• Type: Conference
• Year: 2010
• Where: ICML
• Authors: Xiaojin Zhu, Bryan R. Gibson, Kwang-Sung Jun, Timothy T. Rogers, Joseph Harrison, Chuck Kalish
• Keywords: ICML 2010, Identical People, Machine Learning, Test Items, Test-Item Effect