There is much interest in the way that information about natural scenes is encoded by the visual cortex (V1), whose code needs many more neurons to be effective than a well-designed artificial one; perhaps the V1 code has compensatory advantages, such as energetic ones. Theoretical studies generally model V1 neurons as if they perform arithmetic to the same immense precision as the computers that perform the simulations. The responses of real neurons are imprecise since they are quantised by spike encoding, but could V1 be more robust to quantisation than artificial codes?

We have simulated several model arrays of V1 neurons to see how they might encode the spatial information in 16×16 pixel fragments of photographs of natural scenes. We used a full 256-filter set of the Principal Components (PCA) of such fragments, a set of 256 Gabor-like filters (each like a simple-cell receptive field), and a highly redundant (but more realistic) set of 1364 Gabors. For each set, we calculated how each filter responded to 10000 image fragments, and then how well each fragment could be reconstructed from the responses, using the pseudoinverse of the filter set.

PCA is a complete code, and so reconstruction is perfect, at least to the high arithmetic precision of the computer. To simulate the effects of a simple spike-rate code for response magnitude, we quantised the calculated filter responses before performing the fragment reconstructions. For each degree of quantisation, the 256-Gabor set performed 10-30 times worse than the 256-PCA, but the 1364 Gabor set (with its many more filters) performed slightly better than PCA.For all sets, as the number of quantisation levels increased, reconstruction errors fell. However, once there were more than 50-200 quantisation levels, there was less marked improvement in fragment reconstruction. A real V1 neuron is unlikely to be able to generate 50-200 spikes in the time that an animal takes to identify a visual stimulus (say 50-200 ms), but it has been argued that human contrast coding requires the cooperation of populations of neurons, with a maximal net response of about 180 spikes (Chirimuuta et al. 2003).

To encode image fragments with the same fidelity as PCA, the Gabor sets needed more ‘spikes’. They also required at least as many neurons, so both of the Gabor sets are less metabolically efficient than PCA (Attwell & Laughlin, 2001). V1 is thought to have a Gabor-like code, but its benefit does not seem to be in conferring an energetic advantage.

HB received a Rank Proze Studentship