Emergent emotion via neural computational energy conservation on a humanoid robot

Abstract

This paper presents our initial work on how emotion based behaviors may emerge through computational mechanisms. We hold that in addition to basic emotions such as anger and fear that serves bodily well being of the organism, high level emotions such as boredom and affection have evolved to facilitate low cost brain computations. Higher level of emotions can be considered as affective state of the organism or mood, rather than the reflex-like physiologically triggered emotional responses such as fear and anger. In large and complex brains (e.g. primate brains), the neuronal energy consumption for cognition is non-negligible. We propose that for such organisms computational regulatory mechanisms for decision making give rise to behaviors that can be explained by various emotional states. As a proof of concept for this idea, we envision a robotic cognitive system and a select function that we assign a neural cost for its operation. To be concrete, we use a small humanoid robot platform (Darwin-OP) and implement a neural network (Hopfield Network) that allows the robot to recall learned patterns that it sees through its camera. As a model of neural computational energy consumption, we postulate that a change in the state of a neural unit of the network consumes one unit of (neural) energy. Therefore, the total computational energy consumed is determined by the incoming stimuli. The robot is programmed to avoid high energy consumption by showing aversive behavior when the energy consumption is high. Otherwise, the robot demonstrates engaging behavior. For an external observer these responses may be perceived as robot's having certain emotional (affectional) preference for input stimuli. In this article in addition to robot experiments, we also emphasize the biological support for our proposal and provide detailed exposition of biological background and its relevance for the hypothesis that (certain) emotions may emerge through computational mechanisms.