Abstract
Direction selectivity depends on interactions over both space and time. Second order mechanisms involve such interactions, and previous work by Gaska et al. (1994), Emerson and colleagues (1987, 1992, 1997), and Baker (2001) indicated that direction selective cells were better described by the second order kernel than by the first order kernel. We reexamined this issue by computing first and second order kernels from responses to white noise stimuli for 133 V1 neurons in anesthetized monkeys and cats. From each kernel, we predicted responses to drifting grating stimuli, and compared these predictions to the actual responses. Although first order kernels did not usually provide adequate accounts of responses to gratings, second order kernels were even more inadequate. By a large margin, the first order kernel contributed more than the second order kernel to the predictions of both simple and complex cell responses, based on a percent variance measure. The full predictions never captured more than half of the actual response variance. The second harmonic component of the second order kernel's prediction was weak in all cases, so that the DC component dominated. Alternatives to kernel methods can yield better predictions, and provide models that correspond to the physiology. We derived first order receptive fields for a set of model excitatory and inhibitory inputs using a form of spike-triggered covariance. Although the relation to actual inputs remains untested, the model inputs may shed light on how selectivity is generated. For instance, multiple spatiotemporally oriented inputs can produce non-oriented average maps. The results suggest that quasilinear multi-input models provide insight into the construction of direction selective cells. The models derived here suffer, nonetheless, from the use of white noise stimuli, which deemphasize the low temporal frequencies preferred by cortical neurons.
supported by EY06459 and EY08098