Neural network dynamics of inhibition: Metacontrast masking

Gregory Francis

Purdue University
1364 Psychological Sciences Building
West Lafayette, IN 47907-1364


The dynamic properties of a neural network model of visual perception, called the Boundary Contour System, explains characteristics of metacontrast masking. Computer simulations of the model, with a single set of parameters, demonstrate that it accounts for the key findings of metacontrast masking: strongest masking occurs at positive stimulus onset asynchronies (SOA); masking is weak for negative SOAs; masking effects weaken with spatial separation. The properties of metacontrast masking arise from interactions between positive feedback and lateral inhibition in cortical neural circuits. The model links properties of metacontrast masking with aspects of visual persistence, spatial vision, and cells in visual cortex.

1 Introduction

For nearly a century, researchers interested in psychophysical properties of visual perception have investigated a remarkable phenomenon called metacontrast masking (see Breitmeyer, 1984 for a review). A metacontrast masking display consists of a briefly flashed target (often a filled circle or a bar presented for less than 100 milliseconds) followed by an equally brief masking stimulus (a surrounding annulus or two flanking bars). In such a display the target is perceptually weaker (dimmer) and in some situations subjects fail to perceive the target at all. Perhaps most remarkable, the strongest masking effect often occurs not with simultaneous onset of the target and mask, but at a positive stimulus onset asynchrony (SOA).

The effect of SOA is surprising because if simple lateral inhibition produced this type of masking, then the strongest interactions between target and mask should occur with simultaneous onset. With a positive SOA, it would seem that the information about the target would have moved (to higher visual areas) beyond any influence of the mask. For this reason metacontrast masking is also often called backward masking, thus indicating the apparent ability of the mask to influence percepts of the target by going backward in time. Interestingly, when the mask precedes the target (forward, or paracontrast, masking) there is little masking. Thus, there is a strong temporal asymmetry between the leading and trailing stimuli.

This paper describes the dynamic characteristics of a neural network model of visual perception. Grossberg & Mingolla, (1985a,b) originally built the model, called the Boundary Contour System (BCS), to account for spatial properties of visual perception. Significantly, the same model mechanisms also explain properties of metacontrast masking.

2 Cortical model of boundary information

Grossberg (1994) recently reviewed the model and described its relations to other parts of visual perception. This paper will only describe the model in general terms. The model's functional purpose is to identify the location and orientation of stimulus edges or boundaries. It accomplishes this by feeding a visual image to simple cells, each with a receptive field tuned to changes in luminance at a specific location and orientation. Signals from these cells contribute to other cells that become insensitive to the direction of luminance change, but remain sensitive to orientation and position. These complex cells than feed into a series of cooperative and competitive stages that selects a consistent pattern of cell activations. Figure 1 schematizes the network architecture, where cell icons suggest the shape of receptive fields.

Figure 1: Schematic diagram of the BCS network. Oriented cells that match properties of visual cortex interact in a feedback loop. Solid lines indicate excitation and dashed lines indicate inhibition. The structure of cells schematized here is repeated at every pixel location of the network.

A key component of this selection process is the use of excitatory feedback. The equilibrium response of the network will mostly include cells that send positive feedback to each other (directly or indirectly). Among other properties, this type of feedback allows the network to complete broken contours that occur due to shadows of retinal veins or other types of noise. Thus, excitatory feedback is critical for the network's ability to process spatial information.

The benefits of excitatory feedback for spatial processing comes at a cost of temporal processing. Each cell in the BCS has its own local dynamics involving activation by inputs and passive decay (of the order of simulated milliseconds). However, the excitatory feedback loops dominate the temporal properties of the BCS. When inputs feed into the BCS they trigger reverberatory circuits that are not easily stopped. Simulations in Francis, Grossberg & Mingolla (1994) demonstrate that, if left unchecked, these reverberations can last for hundreds of simulated milliseconds. However, these reverberations do not last indefinitely because internal processes in the network automatically speed disappearance of persisting neural signals. Francis et al. (1994) identified two mechanisms embedded in the BCS design that reset the feedback loop. One mechanism is a type of cortical afterimage that inhibits persisting signals, and the other mechanism is lateral inhibition. The latter plays a dominant role in metacontrast masking.

3 Simulations

The simulations are nearly identical to those in Francis et al. (1994), where simulated stimuli are presented to the network on a 40X40 image plane for a specific duration. The activation of each network cell obeys a differential equation, and the equations are integrated through time to explore the dynamic characteristics of the system. All simulations use the same equations and parameters. The simulations of metacontrast masking make the reasonable assumption that the quality of the visual percept depends on the total duration of target boundaries. For example, longer total durations of target signals should lead to brighter percepts of a bright bar because the development of a brightness percept depends on the presence of boundary signals (Grossberg & Mingolla, 1985a,b).

3.1 Metacontrast and SOA

The primary property of metacontrast masking is that the strongest masking effects occur at positive SOAs. The range of SOAs for maximal masking tends to be around 50-100 milliseconds, but it varies for subjects and stimuli. Figure 2a shows the effects of metacontrast masking for one subject in a study by Growney, Weisstein & Cox (1977). In this study, subjects observed a target (vertical line) with a mask (flanking vertical lines) at varying SOAs and varying edge-to-edge distances. Subjects judged the target's brightness by setting a filter to a standard to produce equivalent brightness percepts; a stronger filter indicates stronger masking. Each curve in Figure 2a shows the masking function for a specific edge-to-edge separation. Within each curve, the strongest masking effect occurs for an SOA of about 90 milliseconds.

Figure 2: Masking is strongest for intermediate SOAs. Paracontrast masking (negative SOAs) is weak. Masking strength decreases with distance. (a) Psychophysical data from Growney et al. (1977). (b) Masking effects on duration of signals in the model. Shorter durations correspond to stronger masking. Note that the y-axis runs in reverse.

The model shows similar nonlinear effects of boundary duration against SOA. While in both short and long SOA simulations the inhibitory signals sent from the mask to the target are the same, activities in the feedback loop are less sensitive to that inhibition for a short SOA because the signals in the excitatory feedback loop are strong. After a long SOA the signals have weakened and are more vulnerable to the inhibition. As a result, increasing the SOA over a limited range, while causing the inhibition to arrive later, decreases the total duration of boundary signals, which is the main metacontrast effect. Figure 2b shows simulation results measuring the duration of a target's boundary signals as a function of SOA and spatial separation from the masks. The results are qualitatively similar to the data in Figure 2a reported by Growney et al. (1973). [Note that the y-axis runs in reverse because shorter boundary durations correspond to dimmer percepts.]

3.2 Paracontrast

Another property from the study of Growney et al. (1977) and obvious in Figure 2a, is that masking effects are weak for negative SOAs, when the mask onset precedes the target onset.

Figure 2b shows that the model has only weak paracontrast masking. This characteristic exists within the model because when the mask precedes the target, the inhibition builds in strength and decays before generation of the target boundaries. As a result, paracontrast masking is weak. The paracontrast masking effects observed in Figure 2b occur because the mask's inhibition can delay the response of cells sensitive to the target. These effects are weak because excitation from the target quickly dominates the fading inhibition from the mask.

3.3 Distance

Finally, the data in Figure 2a from Growney et al. (1977) show that increasing the edge-to-edge separation of the target and mask produces weaker metacontrast masking. Masking effects in this study disappear by three degrees of separation.

In the model, as in the data, the effect of metacontrast masking weakens with spatial separation. In the model the strength of lateral inhibition weakens with distance. Weaker lateral inhibition results in weaker masking. Figure 2b shows that the effects are qualitatively the same as the data in Figure 2a.

4 Conclusions

Francis (1995) describes additional simulations that account for many properties of metacontrast masking. The simulations account for masking with variations of: target luminance, target duration, mask duration, amount of mask contour, and number of masks. Thus, the BCS model is unifying diverse psychophysical data on dynamic and spatial vision.

The dynamic emergent properties used to explain metacontrast masking are consistent with, and depend upon, the BCS's roles in boundary completion, texture segregation, shape-from-shading, brightness perception, 3-D vision and motion processing, among others. Because its properties were orignally built to account for spatial vision, the theory explains not only how metacontrast masking occurs, but why.


April 11, 1996.