Abstract

Our sensory neurons provide measurements of objects, allowing us to categorise them or judge their similarity, but the sensory responses based on these measurements are variable, partly due to neuronal noise. A simple probabilistic classifier for such situations assumes that object likelihoods are Gaussian probability distributions and predicts Euclidean summation of cues in similarity judgments. However, our psychophysical experiments show that similarity judgments to transformed natural visual and auditory stimuli deviates systematically from Euclidean summation. Natural stimuli which differ from an original in two ways are judged slightly less different than Euclidean summation of the two difference cues would predict. This modification of the optimal classifier may be a consequence of the neural events being correlated. Our findings are consistent with a Bayesian classification system that assumes a small correlation (r = 0.12-0.19) between sensory messages. This provides an important link between low level neurophysiology and high level cognitive performance.