Deep learning has been used to help autism therapy actions, as discussed at MIT.
Children with autistic spectrum conditions often have trouble recognizing the emotional states of people around them, and robots have long been used to enable children to imitate their expressions, responding appropriately to sad, cheerful faces, and so on.
Therapy using emotion-expressing robots works much better when the machine is able to interpret the child’s behaviour without problems, and MIT Media Lab researchers have developed a customized type of automatic learning that helps robots estimate each child’s commitment and interest during these interactions, using data that is unique to that child.
Robots trained in human observation could provide more consistent estimates of these behaviours so that in the long run the content of therapy could be customized and make interactions between robots and children with autism more attractive.
The usual AI methods require a lot of data, but they are all similar for each category you learn. Heterogeneity reigns in autism, so normal AI approaches fail. In this case, a human therapist shows pictures of children or cards with different faces intended to represent different emotions, to teach them to recognize expressions of fear, sadness or joy. The therapist then programs the robot to show these same emotions to the child and watches the child interact with the robot.
So far, 35 children between the ages of 3 and 13 have participated, 17 from Japan and 18 from Serbia. During their 35-minute sessions, they reacted in various ways to the robots, from bored and sleepy in some cases to jumping around the room enthusiastically, clapping, and laughing or touching the robot, often treating the robot as if it were a real person.
You can read the complete study in the previous link of news.mit.edu.