MIT researchers have developed a new machine learning model that takes computers one step ahead in interpreting human emotions.
The new model can outperform the traditional systems in detecting small facial variations of humans in comprehending mood. It has been taken through a training of thousands of face images.
The model is said to be able to adapt to a group of people with the usage of ‘a little extra training data.’ It is designed upon the combination of a technique of MoE (a mixture of experts) and model personalization techniques.
According to the paper elaborating the model, the technique combo contributed in detecting highly fine-grained facial expression data from people. This was also the first time the two techniques were combined for effective computing.
Experts in MoE are the neural network models that are trained on a processing task producing one output. Along with this was incorporated a gating network for calculating probabilities on which expert can best detect the moods of unseen entities.
The MoEs were personalized for the training providing video recordings from the publicly available RECOLA database that consists of conversations on a video chat platform which is designed for effective computing applications.
18 video recordings were used from the database, with the experts being trained on nine of those, and put to test across the remaining nine. Furthermore, the experts used a residual network called ResNet for object classification.
After additional tests, it was concluded that the models are potent enough to adapt to populations and individuals with very few data. As per the researchers, data based on skin colors are required for making the models more flexible across diverse populations.
One of the ultimate goals of the researchers for these models is to make them capable enough to help computers and robots to learn from small changing data for naturally detecting how humans feel, in order to better serve us.