A Neural Basis for Perceptual Dynamics

Perceptual stability is ubiquitous in our everyday lives. Objects in the world may look somewhat different as the perceiver’s viewpoint changes, but it is rare that their essential stability is lost and qualitatively different objects are perceived. In this chapter we examine the source of this stability based on the principle that perceptual experience is embodied in the neural activation of ensembles of detectors that respond selectively to the attributes of visual objects. Perceptual stability thereby depends on processes that stabilize neural activation. These include biophysical processes that stabilize the activation of individual neurons, and processes entailing excitatory and inhibitory interactions among ensembles of stimulated detectors that create the "detection instabilities" that ensure perceptual stability for near threshold stimulus attributes. It is shown for stimuli with two possible perceptual states that these stabilization processes are sufficient to account for spontaneous switching between percepts that differ in relative stability, and for the hysteresis observed when attribute values are continually increased or decreased. The responsiveness of the visual system to changes in stimulation has been the focus of psychophysical, neurophysiological, and theoretical analyses of perception. Much less attention has been given to the role of persistence, the effect of the visual system's response to previous visual events (its prior state) on its response to the current visual input. Perceiving an object can facilitate its continued perception when a passing shadow briefly degrades its visibility, when attention is momentarily distracted by another object, when the eyes blink, or when a random fluctuation within the visual system potentially favors an alternative percept. Having perceived an object's shape from one viewpoint can facilitate its continued perception despite changes in viewpoint that distort its retinal projection, potentially creating a non-veridical percept. These examples highlight the importance of the visual system's prior state, not just for perceptual stability, but also for perceptual selection; i.e., for the determination of which among two or more alternatives is realized in perceptual experience. In this essay we discuss three neural properties that form a sufficient basis for a theory of perceptual dynamics that addresses the relationship between persistence, responsiveness to changes in stimulation, and selection. These neural properties are: 1) Individual neurons have the intrinsic ability to stabilize their activation state. 2) Neurons responsive to sensory information (i.e., detectors) are organized into ensembles whose members respond preferentially to different values of the same attribute (e.g., motion direction). Members of such ensembles have overlapping tuning functions; i.e., a detector responding optimally to one stimulus H.S. Hock and G. Schöner 152 value will also respond, though less strongly, to similar attribute values. 3) The activation levels of a detector affects and is affected by nonlinear excitatory and inhibitory interactions with other detectors. On this basis, we examine the persistence of steady-state detector activation despite the presence of random perturbations, the effect of neural stabilization on a detector's response to stimulation, the crucial role of "detection instabilities" in minimizing perceptual instability and uncertainty for near-threshold stimuli, and the importance of differences in the rate-of-change in activation for perceptual selection. Finally, we demonstrate that the signature features of perceptual dynamics, spontaneous switching between percepts differing in relative stability, and hysteresis, follow from the same three neural properties. 1 Perceptual Stability: Natural or Otherwise Natural, everyday percepts are almost invariably monostable. The same percept occurs each time a stimulus is presented. It rarely happens that two qualitatively different percepts are formed for the same stimulus (this would constitute bistability), and the experience of spontaneous switching between alternative percepts is likewise rare. Because everyday experiences of monostability are so pervasive, stability is not always recognized as an important perceptual property. Not so for James Gibson (1966), who attributed the stability of real-world percepts to the tuning of our visual system to unambiguous, invariant properties of stimulation. Although Yuille and Kersten (2005) take a different position, maintaining that natural images are inherently ambiguous, they join Gibson (1966) and others in disdaining the usefulness of artificial stimuli for an understanding of perception in the natural environment. It is arguable, however, that many natural objects are potentially bistable (e.g., bumps and holes), but there is sufficient disambiguating contextual information in the natural environment to over-ride the potential of such objects to exhibit the dynamical behavior that is readily observed in the laboratory. Indeed, it is the exceptional situations accessible in the laboratory that most clearly bring the fundamental indeterminance of perceptual bistability into the domain of phenomenal perception. Our dynamical research has taken place in well-controlled laboratory settings, where we have studied single-element apparent motion (Hock, Kogan & Espinoza, 1997; Hock, Gilroy & Harnett, 2002), displaced targets embedded in noise (Eastman & Hock, 1999), displaced rows of evenly spaced dots (Hock & Balz, 1994; Hock, Balz & Smollon, 1998; Hock, Park & Schöner, 2002), and the motion quartet (which is described below). Single element displacements result in unique motion percepts, and many stimuli with multiple element displacements result in the perception of unique motion patterns. For instance, parallel horizontal motions are perceived for two vertically aligned elements alternating with two vertically aligned elements that are horizontally displaced, as in Figure 1a. This percept uniquely solves the motion correspondence problem; i.e., how visual elements presented during successive points in time are "paired" with respect to the start and end of A Neural Basis for Perceptual Dynamics 153 perceived motion paths (Ullman, 1979). This is the case even though diagonal motions are in principle also possible for this stimulus. That is, despite single element motion being easily perceived for each independently presented diagonal displacement (Figures 1b and 1c), intersecting diagonal motions are never seen when the two diagonal displacements are combined in the same stimulus, as in Figure 1a. Fig. 1 (a) An illustrative apparent motion stimulus for which there is a unique solution to the motion correspondence problem (horizontal motion always is perceived), even though diagonal motions are possible (b and c). (d) The motion quartet, an apparent motion stimulus for which there are two qualitatively different solutions to the motion correspondence problem. For intermediate aspect ratios (the vertical divided by the horizontal distance between the elements), either horizontal or vertical motion is perceived. Horizontal motion predominates for relatively large aspect ratios (e) and vertical motion predominates for relatively small aspect ratios (f). H.S. Hock and G. Schöner 154 In contrast to the stimulus in Figure 1a, the motion quartet is an apparent motion stimulus for which two qualitatively different solutions can be realized in experience; either horizontal or vertical motion is perceived, the proportion of each depending on the aspect ratio of the quartet (Figures 1d-1f). Our research with this bistable stimulus has included psychophysical experiments and the dynamical modeling of spontaneous switching, hysteresis, selective adaptation, and activation-dependent detector interactions (Hock, Kelso & Schöner, 1993; Hock, Schöner & Hochstein, 1996; Hock, Schöner & Voss, 1997; Hock, Schöner & Giese, 2003; Hock, Bukowski, Nichols, Huisman, & Rivera, 2005; Hock & Ploeger, 2006; Nichols, Hock & Schöner, 2006). We currently are studying the effect of neural feedback on the stabilization of global motion patterns for stimuli composed of multiple motion quartets (Hock, Brownlow & Taler, in preparation). It is for stimuli like the motion quartet that it is possible to directly observe the nonlinear mechanisms that bind stimulus specification with the ongoing neural activity resulting from preceding visual events, revealing fundamental properties of the processing mechanisms that are the basis for perception, not just in the laboratory, but in the natural environment as well. The most fundamental of these properties is neural self-stabilization. 2 Neural Stabilization Whether an individual neural detector is activated by a stimulus or not, random events (perturbations) will cause its activation to fluctuate randomly with respect to some steady-state value. However, the variability of these fluctuations does not increase indefinitely over time. Although at first glance this may not be surprising, the "boundedness" of variability reflects a crucial, though often unrecognized feature of neural behavior, namely, that a neuron's activation is actively stabilized. This idea can be made concrete by starting with any activation level for a neuron at any moment in time, and assuming that there is no interaction with other neurons. A random perturbation, if unconstrained, with equal probability will increase or decrease the neuron’s activation. Assume it increases activation. The next and all following random perturbations will again with equal probability increase or decrease activation. Thus, there is nothing that systematically returns the activation from its increased level. Similarly, if an initial perturbation decreases activation, there is nothing that returns the activation from its decreased level. The same logic applies to any activation state generated by perturbations. Over time, states further and further removed from the initial activation state can be reached (e.g., by the chance event of a number of consecutive random increases in activation) and nothing drives the system systematically back from such states. It is intuitively clear, therefore, that the variance of activation would increase indefinitely over time. A formal argument of this kind led to an account for Brownian motion and the increase in time of the uncertainty about the location of a Brownian particle (Einstein, 1905). A Neural Basis for Perceptual Dynamics 155 The essential feature that keeps the variance of random fluctuations bounded is that successive random perturbations do not increase or decrease the neuron's activation level with equal probability. That is, the effects of random perturbations on activation are not unconstrained. When a random perturbation causes a fluctuation in activation, the change is opposed by the neuron's intrinsic ability to stabilize its activation, which reduces the size of the fluctuation. It is because of this resistance to the effects of random perturbations that there is an upper bound to the variance of random fluctuations in activation. The steady-state activation value of a neuron (or population of neurons) that is thus stabilized against the effects of random perturbations is referred to as an attractor. 2.1 The Biophysical Basis of Neural Stabilization The biophysics of individual neurons provides a mechanism for achieving this stabilization of neural activation (Trappenberg, 2002). Specifically, the electrical potential across the membrane that separates the interior of a nerve cell from its inter-cellular environment is kept stable through the mechanisms of osmotic pressure. Ion pumps keep the concentration of different kinds of ions unequal on both sides of the membrane, the resulting flow of ions being in equilibrium when the electrical potential across the membrane just counterbalances the difference in ion concentration. If the equilibrium is perturbed (e.g., by an electrical current injected into the cell), the flow of ions quickly re-establishes the steady-state membrane potential. With a neuron's membrane potential thus stabilized, synaptic input to the neuron increases the potential, increasing the probability that the neuron's activation will be transmitted to other neurons through action potentials traveling down its axon. In our account of neural dynamics (and most other such accounts) the stabilized membrane potential, averaged over local neural populations composed of hundreds or thousands of individual neurons, is sufficient to account for the mapping of psychophysical events onto patterns of neural activation. To be sure, the mathematical relationship between ion flows that stabilize the membrane potential of individual neurons and the stability properties of neural populations is not well understood. Eggert and van Hemmen (2001) have provided one such account, but it is limited by the simplifying assumptions that the constituents of a neural population are both identical in their responsiveness to stimulation and noninteractive. This notwithstanding, it is reasonable to proceed based on the principle that stability properties of neural populations are inherited from the dynamics through which individual neurons stabilize their membrane potential (Jancke, Erlhagen, Dinse, Akhavan, Giese, Steinhage & Schöner, 1999).