Basal ganglia and memory retrieval during delayed match-to-sample and non-match-to-sample tasks

Abstract

Top-down guidance of visual search requires the ability tobias the processing of visual features towards the charac-teristics of the desired object. Whatever form they maytake, these characteristics (that we will call here a templateof the object) have to be transferred from memory to thevisual cortical areas, either from long-term memory orworking memory. High visual areas like the inferotempo-ral cortex, or visuo-mnemonic areas in the medial tempo-ral lobe have been shown to exhibit sustained activitiesduring delayed matching-to-sample (DMS) or delayednon-matching-to-sample (DNMS) tasks, showing theirinvolvement in working memory (WM) processes [1].Particularly, perirhinal cortex (PRh) is involved in objectrecognition, novelty detection and categorization, and itsimpairment leads to severe deficits in DMS and DNMStasks [2].We designed in [3] a computational model of PRh thatwas able to perform multimodal categorization despitepartial object presentation. Its main feature was the incor-poration of dopaminergic modulation of synaptic cur-rents, which allowed the model to exhibit sustainedactivities in clusters of cells representing a given objectunder an optimal level of tonic dopamine (DA) firing. Wealso observed propagation of activity within a cluster,which leads to the interesting property that the visualinformation contained in a cluster can be retrieved troughthe stimulation of a very limited number of cells by tha-lamic afferents. We therefore formulated the hypothesisthat PRh is potential site where visual memory can beretrieved through partial thalamic stimulation in order tobias processing in the ventral stream through feedbackmodulation.Given the fact that PRh receives connections from parts ofthe thalamus that are tonically inhibited by basal ganglia(BG) structures [4], we here suggest that BG can act as con-troller for memory retrieval in PRh, but without any visualdetails about the object to be retrieved. We tested thishypothesis by building a computational model involvingPRh, BG structures, thalamus and a prefrontal cortex(PFC) area that performs static working memory in amanner functionally similar to [5]. We applied this modelto interlaced DMS and DNMS tasks, where one of twoobjects (A or B) is first presented to the system, followedby a task cue indicating whether the system should per-form DMS or DNMS, then both A and B objects. If the sys-tem selects the correct object in PRh, it receives aprobabilistic reward that elicits a phasic DA burst.The model is composed of competing mean firing-rateneurons that interacts through connections havingweights evolve through an homeostatic Hebbian learningrule. Both PRh and PFC areas project on the input struc-ture of BG, the nucleus accumbens (NAcc), which in turnproject to the globus pallidus (GPi) that tonically inhibitsthe thalamus. Selective inhibition of a GPi cell allows itsefferent thalamic cells to exhibit activity and thereforestimulate the corresponding cluster in PRh. The DA cellsslowly learn to predict reward associated to NAcc patterns