Population coding of natural video stimuli in macaque V1

Although it is widely accepted that neurons in cortex encode information about the external world at the population level, little is still known on the structure of neuronal population codes under natural stimulation conditions. In particular, it is not known whether correlations between neurons are important for encoding information. To characterize the cortical population code underlying visual function, we recorded extracellularly the simultaneous activity of neuronal populations in primary visual cortex (V1) of an anaesthetized macaque, while the animal was viewing a video of monkeys behaving in their cages. Neuronal activity was recorded with an electrode array with approx. 1.5 mm spacing. We could record action potentials from a small cluster of neurons at each of 6 different electrodes. We computed, from 30 repeated presentations of the same video, the probability of each neuron firing in response to each part of the video. From this probability distribution of spike patterns at different times of the video, we computed the amount of Shannon’s Mutual Information I that each neuronal cluster conveyed about the sequence of visual stimuli (methods similar to that of de Ruyter et al, Science 1997). We found that each individual cluster conveyed on average 6 bits/sec of information. In addition, to understand how information from different cells is combined together, we computed the information about the video that can be extracted by observing the simultaneous activity over a small population of neural clusters. The size of the population was varied from 2 to 4, and we took the average of the information conveyed by each subpopulation with a given size. We found that the average information increased linearly with the population size. This suggests that information is conveyed at the population level, and that each neuronal cluster carries fully independent information about the visual stimuli. Then, we addressed whether correlations are an important part of the neural population code. If this was the case, we would expect that a downstream neural system decoding the V1 population activity would lose a large amount of information when not paying attention to cross-correlations between neurons. The importance of correlations in decoding can be formalized in information theoretic terms by computing ΔI, the amount of information that is lost by ignoring information when decoding the population activity (Latham and Nirenberg, J. Neurosci. 2005). We found that ΔI was very small, of the order of 1 for population size in the range 2-4. Thus cross-correlations between cortical neurons were not important for transmitting information about natural stimuli. In conclusion, the above results suggest that V1 represents natural visual stimuli through a distributed population code that combines independent information coming from different neurons without relying on correlations