Rodents perform extremely well in navigating complex environments like dark sewers while looking for food or fleeing from predators. The role of the hippocampus in such spatial navigation efforts has interested researchers from the neuroscience domain for decades: in 1971 so-called ‘place cells’ had been discovered in the rat’s hippocampus. While these cell-types allow the animals to swiftly self-localize in a given environment (cognitive mapping), rodents also seem to rely on grid cells (discovered 2005 in the rat’s hippocampus) for path planning and spatial navigation. In addition, recent publications postulate a relationship between theta waves generated in the hippocampus and decision making in spatial navigation tasks.
Constituting neuroscientific breakthroughs at their time, the above findings reveal some basic traits of the rat’s hippocampus, yet are by far not sufficient to unravel the fundamental working principle of the hippocampal formation. To overcome this caveat, we step away from common wet experiments that are restricted to assess small fractions of neural tissue in the rodent’s brain: we aim at the construction of simulated rodents whose spatial navigation capabilities are driven by an artificial neural network that emulates hippocampal functionality. We will base construction of this network on anatomical findings, and eventually create an artificial rodent by embodying our model of the hippocampus in a virtual ePuck robot that explores complex environments in virtual reality.
Using reinforcement learning techniques, the synthetic rodent is meant to learn from experiences in these virtual environments, and is also intended to acquire knowledge from observations of rodents in different real-world scenarios. With sufficient training, we expect our model to closely mimic natural behavior, and to evolve into a synthetic counterpart of the biological hippocampus found in rodents. Yet, unlike the latter, our synthetic hippocampus is no longer a ‘black box’: it provides unrestricted access to its internal neural circuitry, and thus allows for profound understanding of biological navigation paradigms. Even more, our model can serve as predictable laboratory animal in other experimental trials (e.g., extinction learning studies) that likely involve activation of the hippocampal formation.