 |
In general, the visual system provides an accurate means of measuring the orientation of visual contours such as lines and edges (see Howard and Templeton, 1966 for a review). However, contours presented close together or one after the other in the same location can interact, causing distortions in their apparent orientations. When the lines are presented simultaneously, this effect is known as the tilt illusion, when they are presented successively, it is known as the tilt aftereffect. This thesis will focus on the tilt aftereffect, but since the tilt illusion and aftereffect are widely held to have closely related causes, the tilt illusion will be discussed as well.
The prevailing theory for these effects attributes them to lateral interactions between orientation-specific feature-detectors in the primary visual cortex (Tolhurst and Thompson, 1975). The inhibitory connection strengths between activated neurons are believed to increase temporarily while the subject focuses on an input pattern, causing changes in the perception of subsequent orientations. This occurs because the detectors are broadly tuned, and detectors for neighboring orientations also adapt somewhat (Hubel and Wiesel, 1968). When a subsequent line of a slightly different orientation is presented, the most strongly responding units are now the ones with orientation preferences further from the adapting line, resulting in a change in the angle perceived.
Although the fundamentals of the theory were proposed in the 1970s, it
has only recently become computationally feasible to test it in a
detailed model of cortical function. A Hebbian![[*]](foot_motif.gif)
self-organizing
process (the Receptive-Field Laterally Interconnected Synergetically
Self-Organizing Map, or RF-LISSOM; Miikkulainen et al. 1997;
Sirosh and Miikkulainen 1997; Sirosh et al. 1996) has
been shown to develop feature detectors and specific lateral
connections that could produce such illusions and aftereffects. The
RF-LISSOM model gives rise to anatomical and functional
characteristics of the cortex such as topographic maps, ocular
dominance, orientation, and size preference columns, and patterned
lateral connections between them. Although other models exist that
explain how the feature-detectors and afferent connections could
develop by input-driven self-organization, RF-LISSOM is the only model
that also shows how the lateral connections can self-organize as an
integral part of the process. The laterally connected model has also
been shown to account for many of the dynamic aspects of the visual
cortex, such as reorganization following retinal and cortical lesions
(Miikkulainen et al. 1997;
Sirosh and Miikkulainen 1994b; Sirosh et al. 1996).
The current work is a first study of the functional behavior of the model, specifically the response to stimuli similar to those known to cause the TAE in humans. Because RF-LISSOM is a computational model, it can demonstrate phenomena in high detail that are difficult to measure experimentally, thus presenting a view of the cortex that is otherwise not available. The results suggest that tilt aftereffects are not flaws in an otherwise well-designed system, but an unavoidable result of a self-organizing process that aims at producing an efficient, sparse encoding of the input through decorrelation (as proposed by Barlow 1990; see also Dong 1994, 1996; Field 1994; Sirosh et al. 1996).
The rest of this thesis is organized as follows. Chapter 2 is a survey of related and previous work in the study of tilt aftereffects and early vision in general. Chapter 3 explains the RF-LISSOM system in detail, including the network architecture, activity calculation, and connection weight learning mechanisms. It also presents an overview of previous results with the RF-LISSOM model, and evaluates the biological plausibility of the model. Chapter 4 describes how the realistic cortical orientation map used in the aftereffect simulations was trained. Detailed explanations of the experimental settings such as the training schedule and training parameter values are given in this chapter. The resulting map is compared to anatomical and physiological data from humans and other mammals. Chapter 5 describes the aftereffect experiments and results using this orientation map, and demonstrates that the model closely reproduces the psychophysical data for the tilt aftereffect in humans. Chapter 6 further relates the results of the model to the human data, and speculates on the details of the biological mechanisms causing the observed effects. Some directions for future work are suggested, including an examination of tilt illusions, extensions that may be needed for the model to account for behavior at low contrasts, and studies of aftereffects in other modalities. Chapter 7 summarizes the major conclusions from this study. It is argued that this thesis presents the first detailed and convincing computational explanation for the tilt aftereffect, and that it does so within a very general and biologically-plausible self-organizing model of the afferent and lateral connections within the cortex.