A Federated Affective Computing Framework To Learn From Small Data

The field of Affective Computing has for the past decade transformed into one of the most vibrant areas of research under the domain of Artificial Intelligence. However, the development of large-scale affect-enabled applications has often been held back due to issues relating to the data required to train models for affect recognition. This data often does not exist or, if it exists, has its access restricted due to privacy-related issues. We propose introducing concepts such as Federated Learning to the area of Affective Computing to overcome these shortcomings. Federated Learning will allow affect-enabled platforms to source data from many separate agents, helping train models faster while preserving privacy. We introduce the Federated Affective Computing framework, a Composite AI framework, that improves performance of traditional affective computing approaches. More importantly, it aims to the creation of large-scale affect-enabled applications feasible, even in scenarios where limited data is available. From evaluation that has been carried out on a Federated Affective Computing case study, we prove that the creation of intelligent models is made possible even when large enough bodies of data to train such models do not exist. We also show that we make possible the process of "continuous learning" while reducing the risk of catastrophic forgetting.