Spatial learning and navigation in neuro-mimetic systems : modeling the rat hippocampus

摘要

Intelligent behavior in complex environments involves many basic capabilities. In particular, to achieve cognitive navigation, both biological organisms and artificial agents need to be endowed with spatial learning capacities, that is, they must be able to locate themselves within the environment, to perform complex target-directed behavior (e.g., to reach feeder locations), and to avoid penalties (e.g., collisions with obstacles). Neurophysiological and behavioral experiments on rodents suggest that the Hippocampus is the structure of the brain that provides a neural basis for space coding. Indeed, hippocampal place cells exhibit an explicit representation of the animal's location within the environment. In addition, neurons encoding the orientation of the rat's head in the azimuthal plane have been recorded from the hippocampal formation, i.e., head-direction cells. Place coding and directional sense are crucial capabilities for solving spatial learning tasks. This endorses the hypothesis that the hippocampus plays a functional role in rodent navigation, and that it supports spatial cognition and spatial behavior. The overall objective of this work is to study the neurophysiological mechanisms underlying rodent spatial behavior, and to emulate these processes to design an autonomous navigating agent. In particular, we focus on cognitive navigation and we develop a neuro-mimetic system modeling biological spatial learning. We address the two following questions: (i) How do rodents establish a space code of their environment based on available sensory inputs and on their continuous interaction with the world? We put forward a modular neural architecture based on the functional properties as well as the anatomical interconnections of the brain regions involved in space representation. We model hippocampal place cells and head-direction cells as two neural populations. These two systems are strongly coupled and interact with each other to form a unitary space learning system. Allothetic (e.g., visual and auditory signals) and idiothetic inputs (e.g., vestibular and proprioceptive signals) are combined to establish stable place and direction codes. Unsupervised Hebbian learning is employed to allow the agent to build its worldview based on its own experience. (ii) How can cognitive navigation be accomplished based on the above spatial knowledge? The hippocampal space representation model is used as a basis for achieving target-oriented behavior. Hippocampal place cells drive an extra-hippocampal population of action neurons. Synaptic efficacy between place cells and action cells is modified as a function of dopaminergic target-related reward signals. This results in an ensemble action cell activity that provides goal-directed navigation. Biologically inspired solutions have been shown to be a useful methodology for developing flexible and self-contained artificial navigation systems. The experimental evaluation of the model has been done by implementing it on a mobile robot. We emphasize the importance of continuous interaction between the robot and the environment. This results in an incremental and dynamic development of the navigation system, and enables the robot to adapt its lifelong behavior according to situations that it has never experienced before. The present study has produced some insights in the neurophysiological processes involved in animal spatial cognition (i.e., functional and anatomical predictions). From a robotic viewpoint, we have endowed an artificial agent with animal-like exploration and self-localization capabilities, which allows the robot to accomplish effective target-oriented navigation exploiting its interaction with the world.