Attractor dynamics in morph-shape working memory in macaque prefrontal cortex

The neural basis of working memory has been an important topic in current research. The 'bump attractor' model describes a network where structured recurrent excitation and broader feedback inhibition maintain information via persistent neural population codes. This model successfully explains behavioral variations and many neural activity features. However, while the neural basis of working memory in nonhuman primates has been studied extensively with spatial working memory tasks, the neural mechanisms of object working memory remain neglected. We thus trained two macaque monkeys to perform a delayed match-to-sample task using morphed object silhouettes, with five levels (0%-40% morphing with 10% steps between neighboring difficulty levels) of similarity between the target and the distractor. In 22 recording sessions, monkeys were able to perform the task well above chance level for all morphing axes we used (mean performance was 86%, 83%, 77%, 66%, and 54% for 0%, 10%, 20%, 30%, and 40% morphing respectively). We used multi-contact linear probes to record single-neuron activity from the prefrontal cortex as monkeys performed the task. Out of 423 cells recorded, we found 42 that were selective to shape during the delay period. In these cells, we analyzed ones (n=28) with enough error trials when presented with their selective shape pair. We found that decreased persistent activity (comparing average firing rates in a 700ms window), was present in correct compared to error trials following the presentation of a preferred stimulus, paired t-test,t(27)=2.38, p=0.025). Conversely, increased persistent activity following the presentation of a non-preferred stimulus was more likely to result in errors (paired t-test,t(27)=2.08, p=0.047). These results are consistent with the idea that object working memory is being maintained in the prefrontal cortex and is predictive of performance in shape working memory tasks, just as prior research has already demonstrated in spatial tasks.