Encoding of natural scene statistics in the primary visual cortex of the mouse

Abstract

The visual system has evolved to process ecologically relevant information in the organism’s natural environment, and thus it is believed to have adapted to its statistical properties. The most informative components of natural stimuli lie in their higher order statistical structure. If the primary visual cortex has indeed adapted to this higher order structure—as has been posited by theoretical studies over the last 20 years—neural responses to stimuli which differ in their statistical structure from natural scenes should exhibit pronounced deviations from responses to natural scenes. Theoretical studies argue for a sparse code for natural scenes, where only a few neurons need to be active simultaneously in order to encode visual information. However, it has been difficult to assess the sparseness of the neural representation directly and measure the ‘population sparseness’ in neural populations. Here we use 3D random access and conventional 2D two-photon imaging in mice to record populations of hundreds of neurons while presenting natural movies and movies where the higher order structure had been removed (phase scrambled). This technique allows assessing directly how sparse the representation of natural scenes in V1 really is and how this impacts the functional properties of the population code. First, we show that a decoder trying to discriminate between neural responses to different movie segments performs better for natural movies than for phase scrambled ones (nearest-neighbor classifier). Second, we show that this decoding accuracy improvement could be mainly explained through an increase in the sparseness of the neuronal …