Data from: Constructing concepts without feedback: An empirical investigation of how relational information affects multidimensional concept completion behavior in an unsupervised task

The ability of humans to intentionally learn, without feedback,
unidimensional stimulus relations in categorization tasks has been
empirically established over the past two decades.  However,
whether observers can learn more complex multidimensional stimulus
relations across these unsupervised tasks has not yet been
determined.  We demonstrate across an unsupervised concept
completion experiment that the failure to observe multidimensional
learning in previous experiments may be attributable to factors such as
increased stimulus or task complexity.  We posit that concept
completion is related to category learning in that it reveals the
underlying tendencies that are associated with some categories being
easier to learn than others.  In our experiments, we found
observers readily learned to complete a two-dimensional exclusive-or
concept, evidenced by an increase in object selection as the task
progressed with a decrease in choice response times. We also found that
observers readily learned to complete, as evidenced by similar patterns in
object selection and response time behavior, a more complex
three-dimensional stimulus relation that has empirically been associated
with large amounts of categorization errors in related supervised
classification tasks.  Accordingly, we tested two existing formal
models to determine their ability to account for our observations: namely,
the Simplicity Model and the Generalized Representational Information
Theory (GRIT) basic measure. We show how relational information processing
as expounded in GRIT accounts for the observed completion
behavior.  Overall, our findings show how people gravitate, in a
gradual and composite fashion, towards minimizing the perceived complexity
of categories as much as possible.

在过去二十年中，人类无需反馈即可自主习得分类任务中单维刺激关系的能力，已通过实证研究得到证实。然而，观察者能否在这类无监督任务中习得更为复杂的多维刺激关系，目前尚无定论。我们通过一项无监督概念补全实验证实，此前实验未能观测到多维学习现象，可能源于刺激或任务复杂度提升等因素。我们认为，概念补全与分类学习存在关联，因其能够揭示与部分类别更易习得相关的内在倾向。在本实验中，我们发现观察者可轻松习得二维异或（exclusive-or）概念并完成补全，具体表现为：随着任务推进，对象选择次数增加，同时选择反应时缩短。此外，我们还发现观察者可轻松习得并补全更为复杂的三维刺激关系——该关系在相关监督分类任务中已被实证表明会引发大量分类错误——其证据为对象选择与反应时行为呈现出相似的模式。据此，我们测试了两种现有正式模型，以验证其能否解释本实验的观测结果：即简约模型（Simplicity Model）与广义表征信息理论（Generalized Representational Information Theory, GRIT）基础度量。我们阐释了GRIT所阐述的关系信息加工过程，如何解释本次实验中观测到的补全行为。总体而言，我们的研究结果表明，人类会以渐进且综合的方式，尽可能降低对类别的感知复杂度。