personalized machine learning


Imagine a robot assisting a therapist in teaching children with autism to recognize behavioural cues of joy, excitment, sadness and anxiety of neurotypical individuals so that they can establish better social interactions when they meet in everyday situations. To do this, the robot also needs to be able to read social signals of the children on the autism spectrum, usually from their facial expressions, body movements and voice (as most neurotypicals do), but also other parameters such as heart-rate, skin conductance, etc. However, this is difficult! Remember the famous adage: “If you’ve met one person with autism, you’ve met one person with autism.”? Then you can also imagine how difficult it is for a robot to read social signals of children with autism. What if there was a way for the robot to learn individual differences in children with autism and personalize his interpretations to each child? Not only this would help the robot to become a better “teacher” but could also teach us how each child is different from his/her peers, and what we can do to teach the neurotypicals to “adapt” to persons with autism. Wouldn’t that make everyone happier?

Personalized ML models and applications that can efficiently leverage big data and adapt to each individual, especially in smart healthcare and medical applications, are the foundation of the next generation of machine learning from human data. Towards this end, together with Prof. Roz Picard, the founder of the Affective Computing field, I have introduced Personalized Machine Learning (PML) - a novel shift from the traditional machine learning paradigm. Instead of using 'one-size-fits-all' machine learning models, the PML learning algorithms are designed to provide optimal performance (e.g., accuracy of mood prediction) for each individual in target (risk) group - and not only the average group outcome. So far, I have demonstrated, together with my colleagues, the effectiveness of PML on a number of human-data applications, including personalized monitoring of Alzheimer's disease progression, pain and affect estimation from facial expressions, mood prediction, and robot perception of affect and engagement of children with autism, among others.

To teach this new modeling paradigm, in 2017, I and Prof. Roz Picard created a graduate course at MIT Media Lab - MAS.S61: Personalized Machine Learning (PML).

For the papers on the topic, check my google scholar+Personalized

More details and a book on PML are coming soon ... stay tuned!