Similarities in Face Recognition between Deep Learning and Autism Spectrum Disorders

It is acknowledged that FaceNet, a deep learning-based face-recognition algorithm, failed to replicate certain features of the uncanny valley owing to the large difference between the evaluations by people and FaceNet for specific face images. For the images that FaceNet rated as highly human-like, localized attention to the lower part of the face (e.g., the mouth and chin) functioned as the basis for assessment. Such localized attention is considered to be part of the characteristics of individuals with autism spectrum disorder (ASD). This study investigated the similarity in face recognition between the ratings by FaceNet and those by individuals with ASD. Regression analyses were conducted with the ratings by FaceNet as the dependent variable and those by typically developing (TD) individuals or individuals with ASD as the independent one. The ASD group provided a better explanation of the FaceNet evaluation. These results indicate similarities between FaceNet and individuals with ASD.