New machine learning approaches for continuous affective recognition from physiological signals.

BUGNON, LEANDRO ARIEL

Buenos Aires

Conferencia; Machine Learning Summer School 2018; 2018

Organización del MLSS y fundación Sadosky

Exploiting emotion cues in human-machine interaction has been of growing interest in research community and industry. Current advances in wearable technology, such as heart rate monitoring wristbands, provide new non-invasive, practical and private ways of inferring affective states from physiological changes. These signals can be acquired with relative ease, but the task of labeling spontaneous and continuous affects require several efforts. Thus, we proposed new machine learning approaches to improve continuous affect recognition rate. The first method is a new supervised method based on self-organizing maps (sSOM), which clusters physiological features and affective labels to predict the unlabeled recordings [1]. The second method is based on Extreme Learning Machines (ELM), a type of neural network with a randomly generated hidden layer. The methods are evaluated in a semi-supervised task, using a training set with labeled and unlabeled samples. To this end, sSOM application is straightforward, as feature relations are taken into account to cluster the samples besides unavailable labels. Different semi-supervised machine learning approaches are compared with ELM and sSOM. Methods are validated with the publicly available RECOLA dataset. It consists of physiological registers of 18 participants having natural conversations. Expressed emotions are tagged frame-by-frame by six external raters. Current results indicate an improved performance compared with state-of-the-art methods in the supervised-only case. The semi-supervised approach shows promising results for the task when labels are not available, which could ease the development of affective computing applications in the real world.