Active exploration and working memory synaptic plasticity shapes goal-directed behavior in curiosity-driven learning

Abstract

The autonomous discovery and learning of robotic goals is a challenging issue to address. In this work, we propose a cognitive architecture that supports the autonomous discovery and learning of goals. To do so, we draw inspiration from neuroscience by modeling several brain processes such as attention and exploration that we articulate with curiosity-based learning. Moreover, we employ variational autoencoders and create projections of the latent spaces to dynamic neural fields through linear scaling. The aim of these projections is to investigate synaptic plasticity by varying a scaling factor. We demonstrate that a low scaling factor supports a random exploration strategy that produces more diverse actions with no tolerance regarding the discovery of similar goals. On the contrary, a sufficiently large scaling factor drives the exploration toward uncertainty reduction, focusing exploration as well as generating similar actions. In our case, we postulate that synaptic plasticity in working memory can be crucial for exploration and the learning of goals.