Early detection of autism is an essential first step toward access to intervention and services, which can improve quality of life and long-term outcomes. Although autism screening questionnaires are useful, they require literacy and have lower accuracy in real-world settings. Thus, there remains a need for feasible, accurate, and scalable methods for directly observing and quantifying early signs of autism. We have validated a digital phenotyping application (SenseToKnow) which can be remotely administered on a smartphone or tablet and uses computer vision and machine learning to accurately detect autism in young children. This presentation will discuss how SenseToKnow and other computational approaches based on machine learning are transforming methods for the early detection of autism.