Spatial learning and navigation in the rat : a biomimetic model

Animals behave in different ways depending on the specific task they are required to solve. In certain cases, if a cue marks the goal location, they can rely on simple stimulusresponse associations. In contrast, other tasks require the animal to be endowed with a representation of space. Such a representation (i.e. cognitive map) allows the animal to locate itself within a known environment and perform complex target-directed behaviour. In order to efficiently perform, the animal not only should be able to exhibit these types of behaviour, but it should be able to select which behaviour is the most appropriate at any given task conditions. Neurophysiological and behavioural experiments provide important information on how such processes may take place in the rodent's brain. Specifically, place- and orientation sensitive cells in the rat Hippocampus have been interpreted as a neural substrate for spatial abilities related to the theory of the cognitive map proposed in the late 1940s by Tolman. Moreover, recent dissociation experiments using selectively located lesions, as well as pharmacological studies have shown that different brain regions may be involved in different types of behaviour. Accordingly, one memory system involving the hippocampus and the ventral striatum would be responsible for cognitive navigation, while navigation based on stimulus-response associations would be mediated by the dorsolateral striatum. Based on these studies, the aim of this work is to develop a neural network model of the spatial abilities of the rat. The model, based on functional properties and anatomical inter-connections of the brain areas involved in spatial learning should be able to establish a distributed representation of space composed of place-sensitive units. Such a representation takes into account both internal and external sensory information, and the model reproduces physiological properties of place cells such as changes in their directional dependence. Moreover, the spatial representation may be used to perform cognitive navigation. Modelled place cells drive an extra-hippocampal population of action-coding cells, allowing the establishment of place-response associations. These associations encoded in synaptic connections between place- and action-cells are modified by means of reinforcement learning. In a similar way, simple sensory input can be used to establish stimulus-response associations. These associations are encoded in a different set of action cells which corresponds to a different neural substrate encoding for non-cognitive navigation strategies (i.e. taxon or praxic). Both cognitive and non-cognitive navigation strategies compete for action control to determine the actual behaviour of the agent. Tests of the performance of the model show that it is able to establish a representation of space, and modelled place cells reproduce some physiological properties of their biological counterparts. Furthermore, the model reproduces goal-bas