Neural Noise- Natural image statistics relate to patterns of response covariability in macaque primary visual cortex (V1) (Nature).1
- The work is foundational for neural decoding, population coding, and computational neuroscience with relevance to visual neuroprosthetics.1 1
Weekly enrichment (2026-07-20)
- The study is “Relating natural image statistics to patterns of response covariability in macaque primary visual cortex” by Amirhossein Farzmahdi, Adam Kohn, and Ruben Coen-Cagli, published in Nature Communications in 2025 (article s41467-025-62086-1).2
- The authors extend a generative model of V1 encoding using a Gaussian Scale Mixture (GSM), in which a global “modulator” variable introduces statistical dependence among latent image features, and adopt the neural sampling hypothesis so that instantaneous neuronal activity represents a sample from the posterior over those features.2
- Under the theory, a neuron’s across-trial mean and variance reflect the mean and the uncertainty of the posterior distribution, so response variability directly encodes uncertainty about latent image features rather than just noise.2
- The model’s central, testable prediction is that increasing the size of an image reduces correlations for neuron pairs with overlapping receptive fields (and similar tuning) but increases correlations for pairs with spatially offset receptive fields.2
- This prediction was confirmed in population responses recorded from anesthetized male macaque V1, where receptive-field distance influenced surround modulation of correlations more strongly than tuning similarity.2
- The broader sampling-based framework (Orbán, Berkes, Fiser, and Lengyel, Neuron 2016) predicts that perceptual uncertainty is encoded by the variability, rather than the average, of cortical responses, accounting for noise, signal, and spontaneous variability structure in V1.3
- A direct precursor (Coen-Cagli and colleagues, Nature Communications 2021) showed that neuronal variability reflects probabilistic inference tuned to natural image statistics, with variability reduced by stimulus onset and higher contrast because of reduced uncertainty.4
- The result implies that patterns of covariability (noise correlations) are a signature of probabilistic scene representations in V1, with relevance for population decoding and the design of visual neuroprosthetics.2