The experiments most difficult to reconcile with the feedback models are the cortical inactivation experiments of Ferster et al. (1996) and of Chung and Ferster (1998). In each of these experiments, cortical activity was severely disrupted throughout the layers with one of two different methods. The amplitude of the spiking responses to visual and electrical stimuli were reduced by 90% or more (Figure 4), yet the width of orientation tuning of the EPSPs remaining in simple cells, assessed using two different types of visual stimuli, was unaffected. These results strongly suggest that the cortical circuit does not actually sharpen orientation tuning beyond what is provided by the thalamic input, other than through application of the spike threshold within individual simple cells (Carandini and Ferster, 2000). Secondly, the experiments suggest that the geniculate input to simple cells constitutes a large fraction of the total input, 35-50%, so that the cortical inputs amplify the thalamic input only by a small factor of 2-3. And they do so in a manner that is independent of orientation.
These results do not rule out a strong role for feedback elsewhere in the visual cortex. The inactivation experiments were limited to simple cells in the cat visual cortex. Some of the experimental evidence for feedback, however, comes from other cell types and other species. Changes of orientation tuning over time, for example, are best documented outside of layer 4 in the monkey visual cortex (Ringach et al., 1997). These layers in the cat visual cortex have more prominent lateral connections than layer 4 and so might be in a better position to participate in feedback circuitry. Feedback may contribute to responses in many ways without completely determining the structure of cortical activity patterns as in the feedback models of orientation tuning.
These feedback models in their simplest form produce orientation tuning that is independent of all other stimulus attributes. Yet orientation tuning width narrows with increasing spatial frequency (Jones et al., 1987, Hammond and Pomfrett, 1990, Vidyasagar and Sigüenza, 1985, Webster and De Valois, 1985). To correct the models, it is possible to separate the cortical circuit into subsets, each with its own preferred spatial frequency and its own pattern of excitatory and inhibitory connections that determine its orientation tuning width. The question arises, however, as to how many different stimulus parameters, and therefore how many different circuits, must be built into the cortex to account completely for the variability in cortical responses. A consequence of this limitation is explored in detail by Carandini and Ringach (1997), who show that the feedback models are unable to distinguish overlapping gratings of different orientations. Despite the presence of two orientations in the stimulus, the models often converge on their standard pattern of activity within a single group of active cells with a single range of orientation preferences, again defined by the geometry of the excitatory and inhibitory connections. Furthermore, adding noise to a stimulus of a single orientation results in spurious responses in cells tuned to the orthogonal orientation. Such behaviors have not been reported in cortical cells. Rigorous tests for such behavior have yet to be carried out, however, and would form an extremely strong test of the feedback models.