Top-down influences on visual processing

Abstract

Re-entrant or feedback pathways between cortical areas carry rich and varied information about behavioural context, including attention, expectation, perceptual tasks, working memory and motor commands. Neurons receiving such inputs effectively function as adaptive processors that are able to assume different functional states according to the task being executed. Recent data suggest that the selection of particular inputs, representing different components of an association field, enable neurons to take on different functional roles. In this Review, we discuss the various top-down influences exerted on the visual cortical pathways and highlight the dynamic nature of the receptive field, which allows neurons to carry information that is relevant to the current perceptual demands.

Figure 1. Feedback pathways carrying top-down information

Processing visual information involves feed forward connections across a hierarchy of cortical areas (represented by the blue arrows) beginning in primary visual cortex (V1), which in turn receives input from the lateral geniculate nucleus (LGN). The feed forward connections extend through a ventral pathway into the temporal lobe and a dorsal pathway into the parietal and prefrontal cortex. Matching these feedforward connections are a series of reciprocal feedback connections (represented by the red arrows), providing descending top-down influences that mediate “reentrant” processing. Feedback is seen in direct cortico-cortical connections (those directed towards V1), in projections from V1 to the LGN, and in interactions between cortical areas mediated by the pulvinar. Information about motor commands, or efference copy, is fed to the sensory apparatus by a pathway involving the superior colliculus (SC), medial dorsal nucleus of the thalamus (MD) and frontal eye fields (FEF). In addition to direct reciprocal connections, for example from V2 to V1, feedback can cascade over a succession of areas, for example PF to FEF to V4 to V2 to V1. As outlined in this review, a diversity of information is conveyed across these pathways, including attention, expectation, perceptual task and efference copy. (Adapted from Gilbert, Figure 25-7B in Principles of Neuroscience, Kandel, Schwartz, Jessell, Siegelbaum and Hudspeth).

Figure 2. Task-dependent changes in neural tuning and information content in V1

Monkeys were trained to perform two different tasks with a visual stimulus consisting of 5 lines – a central line flanked by two collinear and two parallel lines. Each of the pairs of flanking lines were presented in one of 5 offsets relative to the central line fixed in the receptive field of a recorded neuron, forming a total of 25 stimulus conditions. From these stimuli the animals were cued to perform either a 3-line bisection task, based on the relative positions of the 3 parallel lines, or a vernier discrimination task, based on the relative positions of the 3 collinear lines. The bisection task involves judging to which of the two flanking parallel lines the central line is closer, and the vernier task involves judging the direction of offset of the central line relative to the two collinear lines. (a) The tuning of neurons to the offset of the side-flanks was measured when the animal performed either the 3-line bisection task, where the side flank position was relevant to the task (solid red line), or the vernier discrimination task, where the side flank position was irrelevant to the task (dashed black line). The cell shown in this example was more modulated in its response to side flank offset position when the animal performed the 3-line bisection task (difference in response shown in blue). (b) The change in tuning of a V1 cell to the end-flank offset position when the animal performed the vernier discrimination task, where the tuning was relevant to the task (solid red line) versus when it performed the 3-line bisection task, where the tuning was task irrelevant (dashed black line). (c) The difference in tuning for task relevant and task irrelevant conditions was characterized in terms of mutual information, where the population of recorded neurons carried more information relative to side flank tuning (blue ×) or vernier tuning (red +) in the task relevant condition than in the task irrelevant condition. A series of Monte-Carlo simulations where the responses were randomly assigned to the two different tasks are shown in the blue and red clouds, which are located on the diagonal and far from the experimental conditions. (d) The difference in response between the task relevant and task irrelevant conditions arose from the outset of the neurons’ responses, indicating that the cortical state for performing a given task was set in advance of stimulus onset. (from Li et al, 2004 fig 2, 3, 4 and 7).

Figure 3

Neurons in the prefrontal cortex carry out different functions in accordance with task. Top, monkeys were trained to discriminate between “dog” and “cat” categories in a delayed match to sample task as images were morphed from dog to cat prototypes, or between “sports car” and “sedan” categories as imaged were morphed from sports car to sedan prototypes. Bottom, an individual neuron in the prefrontal cortex showed similar responses to images on one side of the category boundary and distinct responses to images on opposite sides of the category boundary. The differential responses during the delay period between dog/cat categories or sports car/sedan categories support the idea of neuronal multitasking. (from Cromer et al, 2010 figure 2).

Figure 4. Learned association generates recall-related activity in area MT

Area MT normally responds to moving stimuli. However, when trained to associate a moving stimulus, a set of dots moving in a particular direction, with a static stimulus, an arrow (top), neurons become activated by the static stimulus. Bottom A, A neuron in area MT responds to and shows directional tuning to both the moving dot stimulus (red) and the static arrow stimulus (blue). B, for this neuron, polar plot showing tuning to direction of movement (red) and to arrow orientation (blue). (from Schlack and Albright, 2007 figure 2).

Figure 5

Task dependent changes in local field potential coherence and noise correlations in area V1. Neurons were recorded with a 96 electrode array in animals trained on the 3-line bisection or vernier discrimination tasks based on the 5 line stimulus (a, b) or on the contour detection task based on a series of collinear line segments embedded in a background of randomly positioned and oriented lines (c). The effective connectivity between cortical sites representing parallel flanks (a) and collinear flanks (b) was measured by calculating the coherence between local field potentials (LFPs) at different frequencies. The graphs in the center column represent LFP-LFP coherence during the response interval from 100 to 500 ms in the task relevant (red) and task irrelevant (black) conditions. Operations involving grouping of parallel sites, 3-line bisection, or of collinear sites, contour detection, give stronger coherence in the task relevant condition. Operations involving segregation of collinear sites, vernier discrimination, produces weaker coherence in the task relevant condition. The difference in coherence in the 3-line bisection and vernier tasks was seen not only during the entire response period but in the interval preceding stimulus presentation, indicating top-down setting of lateral cortical interactions in advance of the appearance of the stimulus. (d), Noise correlations show task dependent differences. Calculated as Fisher information as a function of changes in stimulus bar position for the three task conditions (black, attend-away, green, attention to the receptive field location, red, performing the relevant task at the receptive field location), the V1 network carried more information about the stimulus when the animal performed the task, and roughly equal contributions to the increase in information came from the changes in neuronal tuning (dotted red line) and from the changes in noise correlation (solid red line). (from Ramalingam et al, 2013).

Figure 6

Top-down influences dynamically change effective connectivity within and between cortical areas, allowing neurons to select inputs, and take on functional properties, that are appropriate for the immediate behavioral context. As a result each cortical area and each neuron within that area is an adaptive processor, continuously changing its line label to serve different functions. Right, Long range horizontal connections link distant points in each cortical map, mediating an association field that provides a set of potential interactions. The association field in V1 is represented by the gray cocircular and collinear lines and by the fields of oriented line segments on either side of the central black neuron. The underlying circuit is represented by the long range horizontal connections formed by excitatory neurons (triangles) and disynaptic connections involving inhibitory neurons (circles). Depending on the top-down instruction, different sets of inputs can be gated according to the state of feedback (represented by the green connections coming from higher order cortical areas), so that under different tasks the black neuron may select either the red or blue inputs. Because of the multiple sources of long range inputs coming from within the same cortical area and from many other cortical areas, and because these influences can cascade over multiple nodes, each neuron effectively becomes a microcosm of nearly the entire brain. Left, multiple layers of such interactions operate across the entire visual pathway, each cortical area containing its own gate-able association field, and top-down interactions cascade across the layers (feedforward pathways are represented by the blue connections between cortical “planes” and feedback pathways are represented by the red connections), not just between nearby cortical areas but also by longer range connections that skip over multiple stages (not shown). Each cortical area is represented here as a 2-dimensional network, but because of their laminar structure different layers tend to be responsible for feedforward connections (superficial cortical layers) and feedback connections (deep cortical layers).