A neural mechanism of dynamic gating of task-relevant information by top-down influence in primary visual cortex

Abstract

Visual recognition involves bidirectional information flow, which consists of bottom-up information coding from retina and top-down information coding from higher visual areas. Recent studies have demonstrated the involvement of early visual areas such as primary visual area (V1) in recognition and memory formation. V1 neurons are not passive transformers of sensory inputs but work as adaptive processor, changing their function according to behavioral context. Top-down signals affect tuning property of V1 neurons and contribute to the gating of sensory information relevant to behavior. However, little is known about the neuronal mechanism underlying the gating of task-relevant information in V1. To address this issue, we focus on task-dependent tuning modulations of V1 neurons in two tasks of perceptual learning. We develop a model of the V1, which receives feedforward input from lateral geniculate nucleus and top-down input from a higher visual area. We show here that the change in a balance between excitation and inhibition in V1 connectivity is necessary for gating task-relevant information in V1. The balance change well accounts for the modulations of tuning characteristic and temporal properties of V1 neuronal responses. We also show that the balance change of V1 connectivity is shaped by top-down signals with temporal correlations reflecting the perceptual strategies of the two tasks. We propose a learning mechanism by which synaptic balance is modulated. To conclude, top-down signal changes the synaptic balance between excitation and inhibition in V1 connectivity, enabling early visual area such as V1 to gate context-dependent information under multiple task performances.

Introduction

The dominant view of cortical sensory processing is that as one proceeds along the visual pathway from primary visual cortex (V1) to higher-ordered visual areas, neurons become sensitive to progressively more complex stimuli-‘complexification’. Early areas are thought to analyze simple attributes, such as orientation, direction of movement, and color. But visual processing also involves countercurrent streams of information flow, and the operation between bottom-up information coding from retina and feedback information coding from higher-ordered visual areas (Engel et al., 2001, Gilbert and Li, 2013, Lamme and Roelfsema, 2000). Visual recognition involved in memory, prediction, and attention is processed in distributed cortical and subcortical networks which include the subcortical structures and V1. However, the involvement of early visual areas in cognition has been often underestimated (Gilbert and Sigman, 2007). V1 neurons are not passive transformers of sensory inputs but rather their activity is influenced by attention, task, training, and expectation (Muckli, 2010, Muckli and Petro, 2013, Tong, 2003). V1 receives considerably more feedback than feedforward thalamic afferents and 10 times more axons from V2 than from the lateral geniculate body (LGN) (Budd, 1998). Consequently, information processing in V1 includes top-down influence. However, it is poorly understood how V1 and higher-ordered visual areas concurrently form visual recognition leading to behavior.

Several studies on perceptual learning have demonstrated that V1 neurons can function as adaptive processor, changing their function according to behavioral context, and their responses reflect the demands of the perceptual task being performed (Gilbert and Li, 2013, Gilbert and Sigman, 2007, Li et al., 2004, Li et al., 2006). In the study on two discrimination tasks with five-bar arrays (Li et al., 2004), it was demonstrated that V1 neurons exhibited the changes of tuning curves depending on these tasks. Further study by Ramalingam et al. (2013) showed the differences in the temporal correlations of spikes and local-field potential (LFP) coherence, caused by the two tasks. Their study pointed out the importance of dynamic control in top-down influence, suggesting that the intracortical connections of V1 network could be dynamically organized so as to gate information critical for the relevant tasks. However, it remains unclear how task-relevant information is gated by the dynamic interplay of the connection change of V1 network and the top-down signals reflecting task behavior.

To address this issue, we developed a neural model of the V1, which receives feedforward inputs from LGN and top-down signals from a higher-ordered area such as V2. V1 neurons interact with each other via a balanced connection of excitation and inhibition. We made two assumptions. One is that a synaptic balance between excitation and inhibition in the V1 network is modulated in a different way, depending on which task is relevant. This assumption is based on the evidence that the circuitry in the V1 has long-range horizontal connections linking columns of similar orientation specificity (Gilbert and Wiesel, 1989, Stettler et al., 2002). Another is that top-down signal has a temporal correlation reflecting behavior relevant to each task. The temporal correlation is supported from the experimental results by Ramalingam et al. (2013). We show that a balance between excitation and inhibition in the V1 connectivity is modulated to gate task-relevant information. The balance change well accounts for the modulations of tuning curves and dynamic properties of V1 responses. We also show that the balance change is caused by top-down signals with temporal correlations reflecting the perceptual strategies of the two tasks. We propose a learning mechanism of the balance change. The results provide a neural mechanism of how V1 is associated with visual recognition and behavior.