Modeling familiarity through the combination of Deep Learning and Hebbian training ; Modélisation de la familiarité grâce à la combinaison d'un réseau profond avec un apprentissage Hebbien

editorial reviewed ; Past computational models successfully reproduced familiarity on abstract patterns with continuous neurons. They were inspired by the perirhinal cortex (PRC) which has been shown to be crucial during familiarity recognition. In fact, a small fraction of neurons in the PRC called novelty neurons respond in a stronger manner when new stimuli are presented. Besides, when a stimulus become familiar, its activity in the PRC is reduced compared to novel ones. Based on this framework, we implemented familiarity recognition on natural images with the combination of a Convolutional Neural Network and a two-layers Feedforward Network, the latter uses Hebbian training to learn natural images. To test the abilities of the model, we implemented a forced-choice recognition (FCR) task and performed several simulations. The first simulation showed that the model has a memory capacity of up to 40 images. The second showed that the model seems to exhibit a recency-like effect only when the number of learned images does not exceed its memory capacity. The third consisted of a FCR task only with images from the same semantic category. Results showed that the model performed significantly worse than when targets and lures are from different categories, thus showing an effect of similarity on the performances. These results suggest that the Hebbian model for familiarity learns the global representation of a stimulus by encoding correlations shared by several stimuli. When too many stimuli are presented, the model seems to learn prototypal patterns which couldn’t allow familiarity recognition.