At least four relatively low level physiological models have been proposed to explain the oblique effect. Two of them, which suggest a more robust neural representation for cardinal than for oblique orientations, can be classified as “gain” models. The first suggests that there are more cells (
Mansfield, 1974;
Orban, Vandenbussche, & Vogels, 1984), or more cortical area
(Coppola, White, Fitzpatrick, & Purves, 1998), devoted to horizontal and vertical orientations than to obliques. If this were the case, then oblique adapting and masking gratings would be expected to produce a weaker effect on the intermediately oriented test pattern, by virtue of their diminished neural representation at the cortical site where masks or pre-exposed patterns modify contrast sensitivity. The finding that oblique adapting and masking gratings are not less powerful than horizontal ones provides evidence against this explanation. Measurements of orientation discrimination in the presence of varying amounts of orientation noise also argue against a gain-based explanation
(Heeley, Buchanan-Smith, Cromwell, & Wright, 1997). In a variation on the gain-based model, Dragoi et al. (
Dragoi, Sharma, & Sur, 2000;
Dragoi, Turcu, & Sur, 2001) suggested that a greater cortical area devoted to cardinal orientations makes their responses more stable, or resistant to modification by adaptation to other orientations. In contrast, obliquely tuned cells, which are more likely to be surrounded by cells with different orientation preference, would be more susceptible to adaptation. This model does not explain our finding that a test grating oriented at 22.5° is affected more by adaptation to a 45° grating than by adaptation to a horizontal grating.
The second model proposes that cortical cells tuned to horizontal and vertical orientations are more sensitive than cells tuned to obliques. This explanation accounts for the detection oblique effect, but is not easily reconciled with the observation that angled lines are perceived as tilted toward the nearest oblique
(Lennie, 1971). It also cannot account for experiments that demonstrate the persistence of an oblique effect for vernier acuity when the horizontal and oblique lines are made equally detectable or discriminable
(Saarinen & Levi, 1995). Versions of this scenario, where the orientation-dependent variation in sensitivity arises before the site of pattern adaptation and masking, predict that horizontal stimuli should be more powerful as adapting and masking stimuli than equal contrast obliques. This prediction is contrary to our results.
A third explanation of the oblique effect posits narrower tuning curves for horizontal and vertically tuned cells than for obliques
(Andrews, 1967). This would account for the orientation-discrimination oblique effect because cells tuned to horizontal and vertical orientations would have steeper tuning curves, making them more sensitive to changes in orientation. The gradually sloping tuning curves for obliques would render them less sensitive to changes in orientation
(Regan & Beverley, 1985). Depending on the quantitative parameters of the model, the orientation of minimum angular discrimination performance would not necessarily be 45°
(Regan & Price, 1986). If one imagines that detection is governed by a “winner takes all” process, then the narrowness of the tuning curves (with equal peak sensitivity) should have no effect on the contrast needed to detect a grating. However, if detection is governed by a weighted sum of units stimulated by the test grating, then wider tuning curves for obliques should give them an advantage for detection. This is because more cells tuned to nearby orientations would be stimulated when the test was obliquely oriented. Such a reverse oblique effect for detection has not been observed. (A reverse oblique effect has, however, been shown for two tasks that require the extraction of form from random dot patterns (
Regan & Regan, 2002;
Wilson, Loffler, Wilkinson, & Thistlethwaite, 2001)). Orientation tuning measurements in primate
(De Valois, Yund, & Hepler, 1982) and cat
(Dragoi et al., 2000) do not reveal a variation in tuning curve width with orientation. In our experiment the test was always oriented at 22.5°. The extent to which the two adapting or masking stimuli elevated the threshold would depend on their strength within the neural channel used for detecting 22.5° orientations. By this logic, different tuning curve widths for horizontal and oblique orientations should have no effect on their adapting or masking efficacy on the 22.5° test. Therefore, this model is also unable to account for our findings.
A fourth model proposes that obliquely tuned units contain more intrinsic neural noise than horizontally or vertically tuned units. However, noise-titrated orientation acuity experiments
(Heeley et al., 1997) have demonstrated that differences in noise between cardinal and obliquely tuned units cannot be the cause of the oblique effect.
Although our observers had contrast thresholds that were 0.3 Log units higher for oblique gratings than for horizontal gratings, oblique patterns were not less effective than horizontal ones as adapting or masking stimuli. This result is problematic for the models reviewed above, but it is consistent with the prediction that the decreased visual effectiveness of oblique stimuli arises after the site of pattern adaptation and masking in cortex. To view the experimental results within this hierarchical framework, it is important to review what is known about the anatomical loci of pattern adaptation and masking.
Pre-adaptation to spatial contrast has been shown to produce a tonic hyperpolarization of cells in the cat primary visual cortex, without affecting the stimulus driven modulations of membrane potential. This hyperpolarization makes the cell less likely to reach spike threshold in response to all subsequently presented stimuli in an unselective manner (
Carandini & Ferster, 1997;
Sanchez-Vives, Nowak, & McCormick, 2000). Psychophysical experiments have shown that pattern adaptation produces a decrease in visibility for subsequently presented patterns that is strongest when the test pattern is the same as the adapting pattern (
Blakemore & Campbell, 1969;
Gilinsky, 1968). This additional selective component of pattern adaptation has also been demonstrated in cortical cells (
Carandini, Movshon, & Ferster, 1998;
Movshon & Lennie, 1979), suggesting that in addition to a tonic hyperpolarization, adaptation selectively alters the synaptic weights of the inputs to a cortical cell or modifies the connections between different groups of cells.
Masking has been actively used to study spatial vision for decades, but it has only been recently that detailed physiological models have been proposed to account for masking phenomena (
Carandini et al., 1997;
Foley, 1994; Freeman et al., 2002). A recent series of V1 physiology experiments resulted in the conclusion that the masking effect is generated partly in the LGN and is supplemented by synaptic depression at the thalamocortical synapse
(Freeman et al., 2002). This proposal, that masking originates earlier in visual processing than pattern adaptation, could explain why the adaptation experiment produced a stronger pattern of asymmetry between oblique and horizontal than the masking experiment.
Older models of masking were based on the premise that units respond with a compressive nonlinearity
(Legge & Foley, 1980). The addition of a masking stimulus to a test stimulus drives a given unit into the compressive range, requiring more of the test stimulus to elicit a criterion response. A key feature of such models is that the various units undergo
independent modification of their sensitivities. Renewed interest in contrast gain control sparked a new class of models (
Foley & Chen, 1997;
Watson & Solomon, 1997), which normalize the linear response of each unit by a measure of stimulus energy from a large pool of neurons (
Carandini & Heeger, 1994;
Geisler & Albrecht, 1992;
Heeger, 1992). These models suggest that masking and adaptation are the result of this nonlinear contrast gain control, or normalization, in primary visual cortex.
Based on the assumption of independent sensitivity regulation, an adapting or masking stimulus would elevate the threshold of a test grating if and only if the mask was detected by the same mechanism as the test. With the “normalization pool” scenario, the mask must affect the pooled signal that modulates sensitivity for a particular test. In either case, our results could be due to the neural representation of the oblique stimulus being slightly more powerful than the horizontal one (i.e., a reverse oblique effect) prior to the site of adaptation or masking. This would cause oblique stimuli to excite the test channel more than horizontal stimuli, and therefore result in stronger adaptation and masking.
An alternative possibility is that the strength of the oblique and horizontal stimuli is the same, but the neural channel that detects the 22.5° test is slightly more sensitive to oblique orientations than to horizontal orientations. This would require the orientation tuning curves within the 22.5° channel to be skewed such that the tail would be longer on the oblique side than on the horizontal side, making them insensitive to the major axis but still sensitive to the diagonal (
Figure 1). This model would predict our unexpected result that 45° adapters and masks are more powerful than horizontal ones at raising the threshold of a 22.5° test. The 22.5° channel would contain units that have greater sensitivity to the adapting and masking gratings that are at 45° than to horizontal stimuli, causing greater adapting and masking efficacy for the oblique stimuli.
A model of this sort also produces qualitative predictions that are compatible with several other experimental results. It predicts better orientation discrimination around horizontal and vertical orientations than around oblique orientations. This is because a small change in orientation would produce a greater change in response where the slopes of the tuning curves are the greatest. The skewing of intermediately tuned curves makes those cells, along with those maximally sensitive to 45°, least sensitive to changes in orientation because of their shallower tuning curves for oblique orientations. The model is also compatible with the observation that 22.5° lines are perceptually closer to 45° than 0°
(Lennie, 1971). This is because cells that are most responsive to 22.5° are often excited by orientations that also stimulate more obliquely tuned cells and the similarity in these neural representations could lead to the perceptual similarity of the stimuli. Previous researchers have found greater adapting
(Gilinsky & Mayo, 1971) and masking
(Campbell & Kulikowski, 1966) half-widths for oblique than for horizontal and vertical stimuli: a result that would also be expected if oblique stimuli activate a wider range of orientation channels.
On the other hand, this skewed tuning curve model does not, without further assumptions, account for the decreased detectability of oblique stimuli. It is possible, as we have suggested, that the detection sensitivity losses for oblique stimuli occur at stages of visual processing subsequent to the site of pattern adaptation and masking. Alternatively, if they occur at prior stages, the effect of the asymmetry in tuning at 22.5° must be enough to outweigh them in our experiments. It should also be noted that although our experimental results cast doubt on gain- and sensitivity-based explanations of the oblique effect, they do not directly contradict them. It is possible that gain or sensitivity differences exist, but that they are overshadowed by other mechanisms in our experiments.
The validity of this model has a bearing on the still contentious issue of the role of intracortical connections, as opposed to afferent connections from the LGN, in shaping orientation selectivity (
Ringach, Bredfeldt, Shapley, & Hawken, 2002;
Sompolinsky & Shapley, 1997). Assuming that the distribution of receptive field centers among a cortical unit’s afferents had even or odd symmetry, it could not generate an asymmetrical tuning curve. However, intracortical connections could. To account for our results, such asymmetries must be introduced into the neural representation at or before the site of pattern adaptation and masking.
Prolonged viewing of a grating makes a subsequently viewed grating of similar orientation appear to be tilted away from the adapting grating
(Howard, 1982). This effect, often referred to as the tilt aftereffect or successive tilt contrast, is thought to reflect a skewing of the distribution of activity over orientation-selective cells
(Gilbert & Wiesel, 1990). It is likely that the skewing is produced by the same orientation-selective sensitivity reduction reflected in contrast threshold measures after pattern adaptation. To specifically test the model that channels tuned to tilted orientations are more sensitive to oblique than to vertical stimuli, measurements were made of the tilt aftereffect produced on a tilted test (roughly 22.5° degrees counterclockwise from vertical) by adapting gratings rotated either 15° more obliquely, or 15° more vertically, than the test. The magnitude of the tilt aftereffect was larger with the more oblique adapting grating than with the more vertical adapting grating for three or four subjects tested (
Figure 5). However, the difference in tilt aftereffect magnitude for the two adapting conditions was only statistically significant for observer DM. This asymmetry in the orientation tuning of the tilt aftereffect provides some, if limited, support for the model.
Asymmetry in the orientation selectivity of cells in cat cortex has previously been demonstrated (
Henry, Dreher, & Bishop, 1974;
Rose & Blakemore, 1974), with one study reporting than 60% of cells in cat area 17 showed tuning asymmetries in excess of 20%
(Hammond & Andrews, 1978). Unfortunately, none of these studies reported the relationship between preferred orientation and degree of asymmetry. Allison and Bonds
(Allison & Bonds, 1994) demonstrated that inactivation of the infragranular layers of cat cortex with GABA broadens the orientation tuning of supragranular visual neurons. In most cells, the broadening was asymmetric, suggesting that intracortical inhibition could play a role in producing asymmetric orientation tuning curves. Asymmetries in orientation tuning have not been reported in primate cortex, but it is possible that skewing has not been seen because of a tendency to measure orientation tuning with a small number of orientations and to fit the data with symmetric functions
(Swindale, 1998) or because orientation tuning is now often quantified by the circular variance of a cell’s response to different orientations.