In the context of human interactions, complexity science (CS) provides a way to understand the dynamics that arise from the interplay of different individuals. Recently, the possibility of applying the theory of CS to computer science, has shifted the focus to the research of machine learning methods for studying human behaviors and relational dynamics. Among the various existing AI techniques, facial emotion recognition (FER) has proven to be the best-performing and easiest to use humancomputer interaction (HCI) tool for emotion detection. Despite the numerous existing approaches, the task of FER is not trivial, mainly due to the absence of large enough datasets to train deep learning (DL) models. A widely used solution is transfer learning (TL), which allows a model, pre-trained on enough data, to be used for a specific task where there is much less data available. The aim of our work is to test the effectiveness of TL for FER on an extremely small dataset, to understand which parameters need to be optimised to obtain the best outcomes. The results showed an overall accuracy of 85.54% for our model and revealed the concrete possibility of applying computer science to complex systems typical of the human psyche.
Transfer Learning for Facial Emotion Recognition on Small Datasets
Paolo Barile
;Clara Bassano;Paolo Piciocchi
2024-01-01
Abstract
In the context of human interactions, complexity science (CS) provides a way to understand the dynamics that arise from the interplay of different individuals. Recently, the possibility of applying the theory of CS to computer science, has shifted the focus to the research of machine learning methods for studying human behaviors and relational dynamics. Among the various existing AI techniques, facial emotion recognition (FER) has proven to be the best-performing and easiest to use humancomputer interaction (HCI) tool for emotion detection. Despite the numerous existing approaches, the task of FER is not trivial, mainly due to the absence of large enough datasets to train deep learning (DL) models. A widely used solution is transfer learning (TL), which allows a model, pre-trained on enough data, to be used for a specific task where there is much less data available. The aim of our work is to test the effectiveness of TL for FER on an extremely small dataset, to understand which parameters need to be optimised to obtain the best outcomes. The results showed an overall accuracy of 85.54% for our model and revealed the concrete possibility of applying computer science to complex systems typical of the human psyche.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.