Probing the Inductive Biases of a Gaze Model for Multi-party Interaction
Abstract
The behavior management controls proposed for social robots are mostly designed for highly controlled scenarios. In the real world though, robots have to adapt to new situations, generalizing learned behaviors. To address this adaptation challenge, neural network models with embedding layers could be used. We present here an approach to better understand the inductive biases of our robotic gaze model. It was trained with multimodal features as inputs -- either endogenous or exogenous to the robot. Inductive biases were explored by observing feature representations in the embedding spaces. We found that the model was able to distinguish between the robot speech intentions that either request or provide information. Similarly, pairs of partners seem grouped according to their social behavior (speaking time, gaze). Finally, we checked that these groupings had a real impact on the model's performance. Driving these biases when facing new people should allow to generate adapted behavior.
Fichier principal
Probing_the_Inductive_Biases_of_a_Gaze_Model_for_Multi_party_Interaction.pdf (845.4 Ko)
Télécharger le fichier
Origin : Files produced by the author(s)