Reply to: “Model mimicry limits conclusions about neural tuning and can mistakenly imply unlikely priors”

Abstract

replying to Michael J. Wolff et al. Nature Communications [DOI of MA to be added at proof] (2025)

A key goal of visual neuroscience is to understand how the physical properties of the world are represented by the brain. Efficient coding theory^1,2 states that neural resources allocated to coding environmental features should be proportional to the frequency with which those features are found in nature. We recently found³ a horizontal bias in the neural representation of visual orientation, as measured in humans with electroencephalography (EEG). We then used generative forward modelling⁴, a method of comparing empirical neuroimaging recordings with matched simulated data produced by different population codes, to adjudicate between previously proposed and novel population codes of orientation in the visual cortex. Wolff and Rademaker⁵ replicated our main findings in their own data as well as in a re-analysis of our data: there is a horizontal bias in EEG measurements of orientation. They argue, however, that generative forward modelling has limited utility because it is susceptible to model mimicry, i.e. many different population codes could be responsible for the same pattern of EEG signals. Further, the authors propose an alternative explanation for the horizontal bias observed in EEG, involving an interaction between stimulus vignetting⁶ and a greater spatial representation of the horizontal meridian relative to the vertical meridian^7,8,9. According to Wolff and Rademaker, this explanation is more plausible because it assumes equal representation of cardinal orientations and, in their view, there is little evidence supporting a horizontal bias in prior literature. Here we respond to these alternative explanations.

While we recognise that model mimicry presents a challenge in any inverse problem, we argue, contrary to Wolff and Rademaker, that rational constraints based on established neurophysiology can mitigate this risk. We will first clarify and expand on existing evidence that provides theoretical grounds for expecting a horizontal bias in neural representation, then explain why stimulus vignetting is unable to provide an alternative explanation for our results, and why Wolff and Rademaker’s findings for peripheral stimuli fail to challenge them. Finally, we highlight converging evidence for the horizontal bias obtained across multiple neuroimaging methods.