28 June 2018. Treating a complex disorder like autism is difficult because of the numerous and diverse ways it can be expressed in children with the condition. An engineering lab at Massachusetts Institute of Technology developed techniques for combining robotics with deep learning, a form of artificial intelligence, to personalize autism therapy, with their findings published in yesterday’s issue of the journal Science Robotics.
A team from the Affective Computing group, part of MIT’s Media Lab, led by computer scientist Rosalind Picard is seeking better technological tools to help children with autism spectrum disorder learn and engage with others. Autism spectrum disorder is a collection of neurodevelopmental conditions marked by communication difficulties and impaired social interaction, as well as repetitive and stereotyped patterns of behavior. Some 1 in 59 children have autism spectrum disorder, according to Centers for Disease Control and Prevention in the U.S., with males 4 times more likely to have the disorder than females. Classic autism is considered the most severe form of the syndrome.
This project, led by postdoctoral researcher and first author Ognjen Rudovic, is creating a therapeutic system for autism spectrum disorder with robotics personalized to the behavioral patterns of the individual children. Rudovic — with colleagues from MIT, Chubu University in Japan and University College London — constructed what they call a personalized perception of affect network to enable humanoid robots interacting with children with autism to adapt to the wide variation in behaviors the robots would encounter. In previous work with robotic therapy, the team found wide variations in cultural and individual differences in emotion and engagement in children during robotic therapy sessions.
To develop their personalized perception of affect network, the researchers enlisted 35 children with autism, with about equal numbers from Serbia (18) and Japan (17). The team used video and audio recordings to capture facial expressions, head movements, body movements, and gestures, as well as physiological measurements such as heart rate and body temperature with wrist bands. A panel of human experts coded the data for positive/negative emotional valence, arousal, and engagement.
The team used their data to train commercially-available two foot-tall humanoid robots — Nao devices made by SoftBank — to interact with the children. The Nao’s training used deep machine learning, also known as neural networks, a technique where algorithms are developed and refined from examples, and become more expert and confident as they experience more data. In this case, the data were the coded emotional valence, arousal, and engagement activities displayed by the individual children, as well as their physiological measures, culture, and gender.
Deep learning algorithms, say the researchers, enable the robots to incorporate the data directly into their software. Rudavic notes in a university statement, “Deep learning allows the robot to directly extract the most important information from that data without the need for humans to manually craft those features.”
When the Nao robots interacted with the children, the interactions were evaluated with standard rating scales for autism-related behaviors. The researchers assessed the interactions of children with robots trained personally for each child, compared to robots trained with the complete set of data, but not personalized. The results show out of the 35 children, the personalized robots correctly interpreted the children’s behaviors more often than the generalized trained devices by 5 to 20 percent per child. In addition, the personalized robots interacted more accurately with 21 to 28 of the 35 children, depending on the factors being evaluated.
Picard is founder of Affectiva, a spin-off company from MIT’s Media Lab in Boston developing systems with artificial intelligence to interpret emotional states. The following video shows a Nao robot interacting with one of the children during the training phases of the study.
More from Science & Enterprise:
- 115M Euro Grant for Precision Medicine, Trials in Autism
- Cancer Drug Seen to Reverse Autism Spectrum Symptoms
- FDA Approves Wearable Epilepsy Seizure Detector
- Umbilical Cord Blood Evaluated for Autism Treatment
- Autism App Adapts Facial Analysis Software
* * *
You must be logged in to post a comment.