Social learning is the hallmark of human intelligence, allowing the human to rapidly adapt to new scenarios, learn new tasks and communicate knowledge that can be built on by others in a continuous setting. While new knowledge plays a role in its accumulation, it also supplements and revises previous knowledge to prevent the human from forgetting (most of) previous knowledge.
That being the case, Intelligent systems, whether they are natural or artificial should have the desiderata of being able to quickly adapt to changes in the environment and to quickly learn new skills by leveraging past experiences. However, most current artificial agents cannot learn in this manner: they suffer from catastrophic forgetting. Although the conventional state-of-the-art deep learning models can be trained on a static dataset to impressive performance on a wide variety of tasks, when these networks are trained on dynamic real-world new tasks, previously learned tasks are typically quickly forgotten. Continual learning, by design, can address this very problem by learning continuously in the incremental setting without forgetting previously learned features.
In the setting of social human-robot interaction, social learning more specifically modelling of human non-verbal behaviour such as recognising spontaneous expression or diagnosing the socio-emotional condition in an individual is crucial for adaptive and personalizable social robots. However, the perception model based on conventional deep learning is not able to adapt to such dynamic interactions. They are not able to personalize toward the observed user, failing to adapt to individual characteristics such as distinctive facial features or characteristic attributes or expressivity. To mitigate the limitation of existing deep learning models, there is a need to develop perception models that can continually learn with each user, be sensitive to their expression and dynamically adapt to changing interaction conditions without forgetting what had been learned before.
In the light of the above, a continually learning paradigm is essential for a social robot to personalize and adapt to the socio-emotional behaviour of the user during interaction. This requires social robots to adapt their perception model to each user, accounting for individual differences in non-verbal behaviour and expression. This adaptation needs to adhere to both the individual level for learning to be sensitive towards the individual behaviour of a user and across individuals for generalising its learning to interact with different users.