The primate brain has a limited capacity to process the immense amount of visual information entering the visual system at any given moment. Visual attention provides a solution to this problem by selecting behaviorally relevant information for detailed processing while filtering out distracting information. In macaque monkeys, attention enhances the responses of visual neurons representing the sensory attributes of behaviorally relevant stimuli while suppressing the responses of neurons representing the attributes of irrelevant distractors. During the voluntary allocation of attention, this response modulation is stronger and occurs earlier in the lateral prefrontal cortex relative to upstream striate and extrastriate visual areas.
Furthermore, activation of prefrontal neurons increases or decreases, respectively, the modulation of single-neuron activity in visual cortices. This suggests that the primate LPFC contains a saliency map influencing neuronal activity in visual cortical areas, thus playing an instrumental role in visual selective attention. Evidence for such a saliency map is mainly provided by studies that average the activity of single neurons over multiple repetitions of the same trial condition. This across-trial averaging is performed in order to overcome the substantial amount of trial-to-trial variability in the responses of single neurons.
It is currently thought that the brain averages the activity of many neurons to overcome the variability of neuronal responses…
Coding of Attention by LPFC Ensembles
The orienting of attention is a dynamic process unfolding over a subsecond timescale. However, most neural correlates of visual attention in the prefrontal cortex of nonhuman primates have been obtained by pooling the activity of singleneurons over a series of trials. Although these studies have helped establish a link between attention and single-neuron responses, to be ecologically valid, neural correlates of attention should be derived from single-trial measurements.
Our results show that a machine-learning algorithm using the activity of neuronal ensembles in area can decode the allocation of attention on a single-trial basis and over time windows as low as 100 ms. This result supports the role of LPFC in top-down visual attention by showing that attentional signals originating from this region could modulate visual activity on a timescale coherent with behavior, despite the variability in single-neuron responses.
Coding of Attention and Saccades by LPFC Ensembles
Eye movements orient the retinas toward events that are behaviorally relevant in the environment, reflecting a close relationship between attention and saccades. However, during covert and divided attention tasks, the allocation of attention can be dissociated from saccades, suggesting that these two processes might be subserved by distinct, but related, neuronal subpopulations. Here we showed that distinct ensemble activity patterns signal the allocation of attention and saccade endpoint, providing evidence that these two processes are dissociable at the level of ensembles.
Our results agree with previous studies in area 8A that dissociate the allocation of spatial attention from saccade goal.
Relevance to Cognitive Neural Prosthetics
It has been suggested that decisions, forward estimations, and even learning-related neural signals could be decoded to control a brain-machine interface that would produce behavioral outcomes according to a subject’s intentions and motivations. Our results support this proposal. We found that the focus of attention could be quickly (within 100–400 ms) and reliably decoded using chronic multielectrode array recordings from LPFC and a simple voltage-threshold operation. Moreover, we showed that multiunit activity that excludes spikes from well-isolated single neurons carries sufficient information to accurately decode the allocation of attention.
Previous studies have suggested that the microcircuits within LPFC are very plastic, or dynamically changing as a function of training during a given task. Our results further suggest that, despite such plasticity, visual, attentional, and saccadic representations are encoded within a map that remains stable over time.