Splitter cells inside a random but embodied computational model
Naomi Chaix-Eichel, Naomi from Mnemosyne group INRIA Bordeaux
Abstract: Over the past decades, the hippocampal formation has undergone extensive study leading researchers to identify a vast array of cells with functional properties (place cells, splitter cells,etc). In the present work, we aim at investigating whether the activity of those cells derive from the anatomy of the hippocampal formation and the properties of neuron or derive instead from the actual behavior of the animal. To do so, we simulated an agent navigating inside an 8-shaped track, making alternating choices (T-maze alternating task). We designed a random network, based on the reservoir computing paradigm, that process distance-based sensors and output a direction change (constant speed). Despite its simplicity, the model successfully solved the task while bearing no structural similarity with the hippocampal formation. We subsequently followed the comprehensive and recent review on splitter cells by Duvelle et al., and applied the exact same analysis onto the activity on the cells composing our model. We were able to identify splitter cells (as well as place cells, head direction cells and decision cells) and confirms a significant portion of the theoretical hypotheses of Duvelle et al. regarding splitter cells. Beyond these results, this work strongly suggests that the activity of such cells originates from the actual behavior as opposed to any structural or anatomical origin: any model doing the same task will exhibit the same cell activity. From a broader point of view, this work questions the epistemic role of such cells in our understanding of the hippocampal formation.
