One consequence of the normalization theory is that no effects should be seen when vertical or horizontal adaptation lines are used, since such lines cannot change the supposed vertical and horizontal norms in either direction. More recent experiments have shown that this assumption is invalid, since similar effects occur regardless of the adaptation angle (Mitchell and Muir, 1976). Another likely consequence is that tilt effects should be zero at precisely 45° (the midpoint between the two axes), as Gibson and Radner (1937) found for their subjects and setup. However, subsequent studies have found that the crossover point between direct and indirect effects varies significantly from 45°, over a range of 25° to 50° (Campbell and Maffei, 1971; Mitchell and Muir, 1976; Muir and Over, 1970).
Sutherland (1961) proposed such an explanation originally, and Coltheart (1971) extended it to account for indirect effects. Coltheart noted that some neurons (here called cross-neurons) had been discovered that had two preferred orientations, each orthogonal to the other (Hubel and Wiesel, 1965, 1968). While an adaptation figure is being shown, some cross-neurons (i.e., those having one of their axes matching the figure) would desensitize in the same way as the simpler neurons whose orientation preference matched. However, subsequent figures having orientations nearly 90° from the adaptation figure would activate fewer cross-neurons than otherwise. This would effectively shift the perceived orientation of the test figure towards that of the adaptation figure, since it will result in less activation of neurons having orientation preferences even further from the adaptation figure than the test figure was.
Muir and Over (1970) proposed an alternative theory of indirect effects within this framework. It had been observed that when an oriented figure is presented, neurons with orientation preferences orthogonal to that of the figure actually have less activity than their resting level (Hubel and Wiesel, 1967). Muir and Over (1970) proposed that these neurons would not only not be fatigued after adaptation, they would be actively facilitated, i.e. somehow become more susceptible to future stimulation. This could cause the indirect effect for lines near their orientation preference.
The feature-detector fatigue theory could apparently account for much of the observed phenomena with either extension for indirect effects. However, it is now discounted for a number of reasons, including the following:
Together, the above factors imply that the observed desensitization must somehow depend on the activation of multiple neurons, including inhibitory interneurons, not merely on changes within the neuron itself.
The inhibition theory also correctly accounts for the phenomenon of disinhibition documented for the tilt illusion (Carpenter and Blakemore, 1973) and the tilt aftereffect (Magnussen and Kurtenbach, 1980a). This phenomenon arises for the TAE when multiple adaptation figures with different orientations are presented simultaneously. For a given test line, two figures which each cause direct tilt aftereffects would be expected to cause an even stronger TAE when presented together during adaptation, if the fatigue theory were correct (Blakemore and Carpenter, 1971)). This result follows simply from having more detectors activated, and thus more detectors fatigued. However, such adaptation actually reduces the amount of the TAE (Magnussen and Kurtenbach, 1980a). The reduction might be explained by the inhibition theory -- the two adapting figures would inhibit the response to each other during simultaneous presentation, and thus the total activation of the orientation detectors near the test line would be smaller than it would be when either line is presented alone. This would cause less adaptation than before, so the TAE will be lower as observed.
Furthermore, there is a wealth of evidence for extensive, modifiable lateral connections (Gilbert and Wiesel 1990; Gilbert et al. 1990; Hirsch and Gilbert 1991; McGuire et al. 1991; Weliky and Kandler 1995). However, as discussed in the next chapter, the detailed behavior of these lateral connections is still a matter of some dispute, particularly with respect to whether they are predominantly inhibitory or excitatory. In any case, the currently accepted explanation of the direct TAE is in terms of the inhibitory effects.
Current theories of the TAE are typically vague about exactly what type of mechanism is causing prolonged inhibition to result in the direct TAE. Most seem to assume that some sort of intracellular buildup of inhibitory transmitters occurs in the target cell (Gelbtuch et al., 1986; Masini et al., 1990; Tolhurst and Thompson, 1975). However, the experiments by Vidyasagar (1990) suggest that this simple explanation will not suffice. Vidyasagar was able to excite or inhibit cells in the visual cortex of cats by locally applying excitatory or inhibitory transmitters, but regardless of the duration of the excitation or inhibition the cells did not show adaptation effects when tested with a visual pattern. Yet the cells showed clear signs of adaptation when a visual pattern was used instead, which prompted the neighboring cells to deliver those same transmitters to the target cell. Thus, unless there is a large gap in our knowledge of the neurophysiology of these cells, the adaptation must be occurring elsewhere, not within the target cell. As suggested by Barlow (1990), this thesis will argue that the changes that occur during adaptation are changes in the strength of connections between neurons. These effects will thus only be seen when multiple nearby neurons are activated simultaneously, as they would be for the patterns typically used (cf. Vidyasagar 1990; Wilson and Humanski 1993).
So far, the most prominent explanation is from Wenderoth and Johnstone (1988); Wenderoth et al. (1989), who have proposed that indirect effects arise from lateral inhibitory repulsion from a virtual axis of the testing line. According to this view, the axes of symmetry of an object can contribute to the perception of it, even if they are not visible. A line (or grating) is symmetrical about two axes, one of which is the line itself, the other of which is a line perpendicular to it. The axis nearest the testing figure is assumed to have the greatest repulsive effect on the figure. Thus direct effects would result (as described above) from repulsion from the actual line, and indirect effects would result from repulsion from the virtual axis. The smaller magnitude of the indirect effects would presumably result from the lower perceptual saliency of virtual axes (Wenderoth et al., 1989). Other differences between the indirect and direct effects have been found in a long series of experiments (Wenderoth and Johnstone, 1988; Wenderoth et al., 1989), most of which indicate that indirect effects develop later in time than the direct effects. Wenderoth et al. interpret these differences as evidence that the processing of indirect effects occurs higher in the visual hierarchy, i.e. that the neural substrate for the virtual axis is in areas higher than V1.
The virtual axis explanation does not appear to have been examined critically by other researchers, perhaps because it appears untestable and nearly as ad hoc as that originally proposed by Gibson and Radner (1937). To make it somewhat more concrete, Spivey-Knowlton (1993) has proposed that the virtual axis results from the cross-neurons described by Coltheart (1971). However, as Spivey-Knowlton himself notes, those neurons are present in V1 as well as in the other early areas. Thus if those neurons represent the virtual axis, they would have to be inherently slower to activate than other V1 neurons, since the indirect effects have been shown to have a later onset (Wenderoth and Johnstone, 1988; Wenderoth et al., 1989). Yet De Valois (1982) found no significant differences in the time behavior of cross-neurons relative to other V1 neurons, so these neurons must not represent the substrate for the virtual axis.
Thus the virtual axis remains an unspecified higher-level process, essentially untestable although possible in theory. In contrast, the explanation for the indirect effect that will be described in this thesis is based on entirely local synaptic resource conservation mechanisms within V1. These mechanisms have already been found elsewhere in the cortex, and they follow from general computational principles, while the virtual axis appears to have been invented purely to explain indirect tilt effects.