Purchase this article with an account.
Jiang Mao, Alan Stocker; Orientation perception is based on efficient coding and categorical decoding. Journal of Vision 2021;21(9):2643. doi: https://doi.org/10.1167/jov.21.9.2643.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Perceived stimulus orientations are typically biased away from cardinal orientations. Wei and Stocker (2015) proposed that these biases are the result of a Bayesian inference process constrained by efficient coding, where higher encoding accuracies for cardinal orientations reflect the natural prior of local orientations. Other studies, however, suggested that natural categories for cardinal and oblique orientations also play a role in the perception of orientation (Rosielle & Cooper, 2001; Wakita, 2004; Quinn, 2004). Here, we systematically tested to what degree efficient coding and the notion of category are necessary to provide a quantitatively accurate account of human psychophysical data. Specifically, we tested how well a Bayesian observer model with/without efficient coding and with/without categorical loss can account for the psychophysical orientation estimation data from De Gardelle, Kouider and Sackur (2010). In formulating a categorical loss function, we assumed that the observer considers four natural orientation categories (horizontal, vertical, clockwise and counterclockwise to vertical) relative to a noisy reference. The total loss function then consisted of a weighted combination of a mean squared error loss term and the categorical loss. We find that the model versions without efficient coding cannot explain the direction, magnitude, or the dependency of the estimation bias on sensory noise. The models with efficient coding are all able to predict the repulsive bias. However, the variance pattern predicted by the efficient Bayesian model without categorical loss doesn’t match the data. Only the model version that incorporates both efficient coding and a categorical loss component is able to quantitatively fit the full distribution of orientation estimate in its minute details. In conclusion, we can quantitatively fit the orientation estimation data with a Bayesian observer model that incorporates both efficient coding and a categorical loss function. Both elements are necessary to explain the distortions in orientation perception.
This PDF is available to Subscribers Only