Active machine learning framework for teaching object recognition skills to children with autism

Abstract

Active machine learning (AML) techniques enable a machine learning model to perform better with less labeled training data. In this study, we proposed an AML approach for teaching object recognition skills to children with autism spectrum disorders (ASD) and compared its effects with passive learning (PL). A web and touch-based application was developed for teaching object recognition where objects were grouped according to their categories and difficulty levels. The teaching procedure was based on Applied Behavioral Analysis principles. Five children with mild to moderate levels of ASD participated in the study. An alternating treatments design of single-subject research methods was used. The results indicated that AML was more effective than PL for four out of the five participants. Consequently, they can learn faster with fewer teaching trials that are required to reach a learning criterion.

Keywords:

• Autism spectrum disorder
• Object recognition
• Active machine learning
• ABA
• Technology

Acknowledgements

We would like to thank the educators at the Right to Live Society (RTL) and Palestinian Society for Autism & Rehabilitation (PSAR) for their help in conducting this study. The first author was partially supported by a scholarship from the Turkish government during his thesis studies.

Notes

^1. A model over-fitting can occur when training data is included in test sets. This implies that a model begins to ‘memorize’ training data rather than ‘learning’ to generalize from trend.

^2. In order to protect the children’s identity, fictitious gender matched names are used.