Abstract
Lateral prefrontal cortex (lPFC) neurons are thought to encode working memory (WM) representations of visual space via sustained firing. Neurophysiological studies of WM typically record from individual neurons, thus we lack an understanding of how larger ensembles of simultaneously-recorded neurons represent WM: we do not know if WM representation fidelity is affected by phenomena that are not measurable in single neurons (e.g. noise correlations– rnoise), nor how WM coding properties scale with the size and composition of neuronal ensembles. In order to investigate these questions, we used microelectrode arrays to record from neuronal ensembles in lPFC area 8a of two rhesus macaques while they performed a traditional oculomotor delayed-response task, and assessed the information content of the ensembles using a linear classifier. We found that: 1: The contents of WM (which of 16 locations is being remembered) can be reliably decoded from ensembles of 30-60 units in a 500ms window from a single trial (mean accuracy = 48%; max = 77%; chance = 6.25%). 2: Units that are "non-selective" (ANOVA for remembered location, p ≥ .05) can still increase the information content of an ensemble, entirely by altering the rnoise structure. 3: The most informative ensembles are not necessarily composed of the most informative individual units. 4: Removing rnoise can yield inverse effects on the fidelity of WM representations depending on the size and tuning properties of the ensemble (maximum 5% median decoding accuracy improvement, minimum -7.6% median decrease; p = .002 and p < .001, respectively, Signed rank test). Our results demonstrate the importance of correlated variability in WM representation fidelity in lPFC neuronal ensembles, and that its effects are mediated by the size and composition of the ensemble. However, aggregating single neuron properties does not necessarily predict the properties of larger neuronal ensembles.
Meeting abstract presented at VSS 2016