Figure 1. Example of a dot lattice. Would you say the dots are aligned in strips in the 45 degree, 135 degree, horizontal or vertical orientation?
Perception has proven a tough problem to understand in organisms and to implement in artificial systems. The reasons for this are not only situated in the intertwining of perception with a large knowledge database, but also in the dynamic flexibility with which perceptual systems deal with the information that is picked up at the level of the receptors. Research conducted in the group of prof. Johan Wagemans to adress specifically the latter topic in vision has crystallized around two themes, described below.
Unlike what most people would expect from an intuitive notion of perception, it is not at all a passive absorption of 'what is out there'. If anything, perception is no less than an act. Before one has any idea of one is looking at, an immense computation is taking place. Among other things, the visual system fills in missing parts, and tries to guess what belongs to what in the image it receives through the retinae. The latter phenomenon, the fact that humans (and other animals) are able to perceive separable elements as a whole, also if the elements are spatially discrete, and thus the visual representation of the whole spans across space, is denoted as perceptual grouping. Although many of the factors that play a role in bringing together elements in the visual scene are known since the first half of the 20th century, mainly through the work of a group referred to as the Gestalt psychologists, few details are known as to which mechanisms cause unified percepts. Is it a matter of 'integrative' neurons, sensitive to specific combinatory patterns of activity in lower representations? Or is synchronous activity of groups of neurons sufficient to make them part of a unit? What is the role of long-range connections between groups of neurons with similar receptive field properties? Questions like these, even if they are central to the understanding of how grouping works, are currently unresolved. Sharp psychophysical experiments, neurophysiological recordings, neuroanatomical data, thorough analysis of the behavior of neural systems simulations; all of these are essential ingredients for capturing a full insight into the processes at hand.
Grouping is a computational mechanism that takes advantage of the raw power available to the visual system, implemented in a massively parallel system for calculation. Consider the area called V1, the primary visual cortex; this huge dynamical system (about 100 000 000 000 neurons, with 1170 connections per neuron within a piece of cortex 1mm thick) realizes several phenomena of grouping. The complex and extensive nature of the field of visual grouping forces researchers to focus on a distinct issue, with a limited set of stimuli. The work on visual grouping in the laboratory of experimental psychology of the KULeuven is based on a class of visual stimuli called dot lattices (Figure 1). These are multistable, meaning that there is more than one perceptually feasible solution to how to group the dots. We apply the inherent multistability of stimuli like these to study the weight of and interaction between different grouping cues, such as proximity, collinearity of the elements, or dot polarity. We also have a class of stimuli that we use to investigate how local information influences the grouping odds for the configuration percept, specifically tuned to match known properties of receptive fields in the early stages of visual processing (Gabor lattice, see Figure 2).
Figure 1. Example of a dot lattice. Would you say the dots are aligned in strips in the 45 degree, 135 degree, horizontal or vertical orientation?
Figure 2. Example of a Gabor lattice. The distance between the patches favors another organization than the one supported by local orientation information (through element alignment). Which organization will subjects perceive in this case? What is the neural substrate of the perception of 'togetherness' of elements?
There is still a lot of work in comparing the results of these experiments with findings in broader neuroscience. We have already planned several studies to magnify aspects of previous empirical work, and we could use some help in devising new clarifying experiments. A neural network simulation of grouping processes tailored for the lattices could tell us (and of course the broader scientific community) a lot about how our experimental results fit in with existing or new neural models of brain function in early vision, about the role of horizontal, feedforward- and feedback connections in grouping, and about the plausibility of synchronous firing as a model of unitary perception.
The human nervous system, with its 1012 units with 1015 connections in total, is a huge dynamical system. As a structure that sometimes deals with multistable interpretations, it exhibits, just like other physical systems, some persistence in switching from one state to the other. This feature is known as hysteresis: the tendency of a system to remain in the state it is currently in. It is a phenomenon that is particularly strong in visual perception; this can be illustrated with a Necker cube; the kind of figure that looks like a wireframe cube, of which the threedimensional interpretation has two possible orientations (Figure 3). If you see a Necker cube, you will establish a fairly stable percept that does not change easily. In two subsequent presentations of a Necker cube, it is very likely that your interpretation of the second cube will be the same as the interpretation of the first, even if a priori both interpretations are equally probable. Moreover, if we would gradually change the features of the cube so as to favor the interpretation opposite to the one you had when you first saw the cube, you would tend to stick with your existing percept.
Figure 3. A Necker cube, rotated 45 degrees with respect to its typical illustration. There are two conflicting 3D interpretations for this drawing. How do subsequent interpretations influence each other?
A process that seems to run contrary to hysteresis is adaptation. Adaptation, in this context, refers to the ability of perceptual systems to discard information that remains constant through time. A typical example is the blindness to images that remain fixed on the same spot on the retina (the so-called stabilized images) for a period of time. The visual system, through different mechanisms, 'prefers' to represent changes in the input rather than the entire information; what remains constant is uninteresting. In multistable perception, where two or more interpretations are possible for one image, giving less weight to one interpretation alternative because it occurs repetetively, or over prolonged duration, will cause a higher probability for the other alternative to pop up. This implies that you are more likely to switch interpretation in a second Necker cube if you have been looking at the first for a longer time (and if your percept was stable for the entire presentation).
To summarize: given two possible perceptual states, the hysteresis principle dictates that subsequent states will be the same, while adaptation predicts that a switch will occur, if the first state has been active for a while. Are these predictions irreconcilable? No: experimental data indicate that they work additively in perception of multistable figures. Even if they are both present, depending on the parameter regime (for example the duration or frequency of each stimulus), either adaptation or hysteresis will prove to be the strongest force. Interestingly, it seems to be the case that adaptation also occurs for alternatives that were there, but have not consciously been perceived, while hysteresis has as a condition that a percept has explicitly been experienced. There are still alternative explanations for these data, however.
We have planned an experiment to investigate the relative weights of stimulus duration and repetition in the interpretation of a subsequent multistable stimulus. Also in this research we use dot lattices, because of the straightforwardness of the stimulus when we want to bias our participants to one specific perceptual organization. Here a project can include formalization of the underlying processes, and capturing the experimental paradigm in terms of dynamical systems and the associated concepts for comparison with other dynamical multistable processes, and extracting the mathematical features of the state space of this multistable representation.
Although we can use the help of people with different orientations and skills, projects in this field will particularly fascinate students with an interest in formal modeling, dynamical systems theory, and neural network simulations. This necessarily implies some proficiency in understanding mathematical reasoning and handling equations. Work that is geared more towards psychophysical experiments will have to be supported by good statistics; therefore it is helpful to be acquainted with the way statistical analyses are performed. Computer-literacy and knowledge of a higher-order programming language such as C++ or the MatLab programming environment will greatly reduce practical problems in conducting experiments or simulations. The aforementioned skills are assets; a true prerequisite is the willingness to dive into a research field that is interdisciplinary in nature and requires a creative mind.
If you take up a project in one of the issues described above, you will be guided by the people working in the laboratory of experimental psychology, in the group of prof. Johan Wagemans. More specifically, prof. Johan Wagemans will be the supervisor of the research, while Peter Claessens will be assisting. In case you would like to obtain more information, you can find both at the back, to the left, of the ground floor at the institute for psychology, Tiensestraat 102. You can also contact them through phone or email, Peter Claessens at 016/325930, and email Peter.Claessens@psy.kuleuven.ac.be, prof. Johan Wagemans at 016/325969 and email Johan.Wagemans@psy.kuleuven.ac.be.