Title: A new angle on the role of feedfoward inputs in the generation of orientation selectivity in primary visual cortex
Abstract: Fifty years ago, in this journal, Hubel & Wiesel (1962) published what has become, arguably, one of the most influential articles in systems neuroscience history. Inspired by a systematic comparison of receptive field (RF) structure in thalamus (lateral geniculate nucleus, LGN) and primary visual cortex (V1), they described how V1 neurons could become selective to the orientation of visual contours, a critical first step towards object recognition. Hubel and Wiesel's model considers that simple cells and complex cells, the two main RF classes they originally described in V1, correspond to two successive stages in cortical processing. In the first stage, simple cells could emerge in layer 4 from the convergence of thalamic inputs with RFs arranged along a row in visual space. The spatial structure of the simple RF would thus represent the most parsimonious approach to build orientation detectors directly from LGN cells with circular RFs (Fig. 1A). Two alternative feedforward models of the simple receptive field A, in Hubel and Wiesel's original formulation, simple RFs (top) are generated in layer 4 from the convergent input of On- and Off-centre geniculate neurons with RFs precisely arranged along parallel rows in visual space (bottom). B, the new version of the feedforward model predicts that rather than being aligned in rows, geniculate RFs overlap extensively in visual space (bottom); when linearly combined (middle), they generate population RFs that closely resemble the structure and orientation selectivity of their targets in cortical layer 4 (top). Hubel and Wiesel's hierarchical model, appealing for its simplicity and explanatory power, did for systems neuroscience what Hodgkin and Huxley's model of action potential generation had done for cellular neuroscience a couple of decades previously. Unlike Hodgkin and Huxley, however, Hubel and Wiesel were not at all precise in the mathematical formulation of their proposal and the emergence of orientation selectivity has remained highly controversial ever since. In favour of a hierarchical model, it was recently found that cells with simple RFs are confined to regions that receive direct thalamic inputs. In addition, it was reported that the RFs of LGN afferents are distributed in visual space along the preferred angle of a cortical orientation column, where they overlap the simple cell subregions according to contrast sign and retinotopy (see Hirsch & Martinez (2006) for review). And finally, when the firing of cortical neurons is abolished, the spatial structure of the simple RF remains largely constant and the remaining (presumably thalamic) input to layer 4 simple cells appears to be equally tuned for orientation (see Kara et al. (2002) for discussion). Critiques to the hierarchical model, on the other hand, have mounted over the years. Some studies have even questioned the existence of two discrete cell classes in V1, suggesting that Hubel and Wiesel's simple and complex cells represent the two ends of a continuum of receptive field structures found in all cortical layers. In addition, simple feedforward models cannot easily account for various aspects of cortical responses, such as the contrast invariance of orientation selectivity. Finally, only a small fraction of the excitatory synapses in layer 4 are made by feedforward LGN afferents. These results were used to promote the view that the thalamic input cannot by itself determine the orientation selectivity, let alone other functional response properties, of cells in cortical layer 4. Therefore, several alternatives to the hierarchical model, the so-called feedback models, have emerged which consider that the cortical microcircuit plays the crucial role of amplifying and transforming the LGN input according to context and behavioural state (see Hirsch & Martinez (2006) for review). To contribute even more to confound these two antagonistic paradigms, and the pundits, two recent reports (Jin et al. 2011; Viswanathan et al. 2011, in a recent issue of The Journal of Physiology) now argue against the thalamocortical pattern of excitatory convergence implied in Hubel and Wiesel's model. But, at the same time, they also suggest that orientation tuning at the cortex must already be encoded in the response properties and partial segregation of On and Off cells in the LGN, as was implicit in the original feedforward hypothesis. Jin et al. (2011) have found that the receptive fields of On and Off LGN afferents do not always scatter along the preferred angle of a cortical orientation column. Instead, they overlap extensively and are restricted to a small region of visual space. However, when linearly combined, they generate spatial profiles resembling those of cortical simple cells and have the same preferred orientation as the target cortical column (Fig. 1B). In addition, in a recent issue of The Journal of Physiology, Viswanathan et al. (2011) used a clever variation of the approach of Kara et al. (2002) to isolate thalamic input to the cortex to show that the orientation selectivity seen in the subthreshold activity of simple cells during the silencing of cortical firing could be explained by the orientation biases in the output of single LGN cells, thus reducing the need for excitatory convergence from a long row of thalamic afferents. Therefore, what is it going to be, feedforward or feedback, thalamocortical or corticocortical? These recent reports highlight that distinguishing between these two antagonistic models of cortical organization and function will require a full understanding of how the thalamus transforms the retinal output before it reaches V1. We believe the computations performed at this subcortical nucleus will prove to be fundamental, not only to explain the generation of orientation selectivity in cat V1, but also to account for interspecies differences in the emergence of cortical RFs and maps. If proved to be true, the consequences will be far reaching, and may bring us closer to understanding the emergence of much more complex RF structures and cortical topography in areas downstream of V1. Surprisingly, much can be explained on purely feedforward grounds and without the need to invoke a complex set of developmental wiring rules. A nice new angle on an old riddle.