Dream Net: a privacy preserving continual learning model for face emotion recognition

Marion Mainsant (1) , Miguel Solinas (1) , Marina Reyboz (1) , Christelle Godin (2) , Martial Mermillod (3)

1 LIIM - Laboratoire Intelligence Intégrée Multi-capteurs
2 DACLE-LETI - Département d'Architectures, Conception et Logiciels Embarqués-LETI
3 LPNC - Laboratoire de Psychologie et NeuroCognition

Continual learning is a growing challenge of artificial intelligence. Among algorithms alleviating catastrophic forgetting that have been developed in the past years, only few studies were focused on face emotion recognition. In parallel, the field of emotion recognition raised the ethical issue of privacy preserving. This paper presents Dream Net, a privacy preserving continual learning model for face emotion recognition. Using a pseudo-rehearsal approach, this model alleviates catastrophic forgetting by capturing the mapping function of a trained network without storing examples of the learned knowledge. We evaluated Dream Net on the Fer-2013 database and obtained an average accuracy of 45% ± 2 at the end of incremental learning of all classes compare to 16% ± 0 without any continual learning model.

Keywords: continual learning; incremental learning; pseudo-rehearsal; catastrophic forgetting; privacy; face emotion recognition; replay method; replay method I

Domains: Artificial Intelligence [cs.AI]; Neural and Evolutionary Computing [cs.NE]

Marion Mainsant, Miguel Solinas, Marina Reyboz, Christelle Godin, Martial Mermillod. Dream Net: a privacy preserving continual learning model for face emotion recognition. 2021 9th International Conference on Affective Computing and Intelligent Interaction Workshops and Demos, Sep 2021, Nara - Online, Japan. ⟨cea-03474722⟩