A Neural Dynamic Network Drives an Intentional Agent That Autonomously Learns Beliefs in Continuous Time

Autonomous learning is the ability to form knowledge representations solely through one’s own experience. To autonomously learn, an agent must be able to perceive, act, memorize, plan, and desire; it must be able to form intentional states. We build on a neural process account of intentionality, in which intentional states are stabilized by interactions within populations of neurons that represent perceptual features and movement parameters. Instabilities in such neural dynamics induce sequences of intentional behavior. In this article, we examine the neural process organization required to decide and control when learning takes place, to build the representations that can hold learning data, and to organize the selection of neural substrate to learn the novel patterns. We demonstrate how a neural dynamic network may learn new beliefs about the world from single experiences, may activate and use beliefs to satisfy desires, and may deactivate beliefs when their predictions do not match experience. We illustrate the ideas in a simple toy scenario in which a simulated agent autonomously explores an environment, directs action at objects, and forms beliefs about simple contingencies in this environment. The agent utilizes learned beliefs to satisfy its own fixed desires.