Feature Selection in Category Learning

• In Event: Poster Session 07
  In Poster Session: PS 07 Topic - Attention, Learning, Memory

Abstract

Research examining mechanisms underlying human categorization has reported that when learning novel categories, adults tend to selectively attend to the diagnostic features that can separate different categories, whereas young children who show difficulty focusing on a single relevant dimension tend to allocate attention to multiple features. The current study further investigated mechanisms underlying children and adults’ category learning. By measuring adults’ and 4-year-olds’ accuracy and response time in classification tasks, we aimed to answer the following two questions. First, do adults and children selectively rely on one single deterministic feature when learning new categories? Second, is a single deterministic feature sufficient for adults and children to determine category labels? In Experiment 1, 4-year-olds and adults were trained with categories that had a single deterministically predictive feature and multiple probabilistic features, and they were tested with items varying in the number of visible features. In Experiment 1, participants were tested on four different item types: (1) High Match (items that are similar to training items); (2) Switch items (these items are a mixture of the deterministic feature from one category and the overall similarity from another); (3) D+P (items showing the deterministic feature and one probabilistic feature, with the rest of the features covered); and (4) P+P (items showing two congruent probabilistic features, with the rest of the features covered). The results indicated that with sufficient training, both adults and children relied exclusively on the deterministic feature regardless of overall similarity. Importantly, a deterministic feature was both sufficient and efficient for learning new categories. When the deterministic feature was present, participants were as accurate and fast when classifying items with most probabilistic features missing as when classifying items with all features present. However, when the deterministic feature was inaccessible, their accuracy dropped, and response times slowed. In Experiment 2, to further investigate participants’ attentional optimization and reliance on the single deterministic feature, we included NoD, an additional item type for which only the deterministic feature was covered. The results for adults showed that adding more informative probabilistic features didn’t significantly improve their accuracy. Adults were equally accurate on classifying items with only two congruent probabilistic features present and on classifying items with five congruent probabilistic features present. The finding provided additional evidence that when learning novel categories with a single deterministic feature, adults optimize their attention by selectively attending to the deterministic feature while ignoring other probabilistic features. Although due to COVID-19 pandemic, data collection with children is still underway, we expect that, in contrast to adults, children will not fully optimize attention in the course of learning. As a result, unlike adults, they will exhibit higher categorization accuracy for NoD items than P-P items. More children’s data will be collected in the next five months to test our hypothesis, and data analysis will be completed by the time of the meeting.

Authors

• Mengcun Gao, Ohio State University
  Presenting Author

• Vladimir Sloutsky, Ohio State University
  Non-Presenting Author