Abstract
Behavioural pattern detection experiments have greatly advanced our understanding of the computations performed by the early visual system to extract information from the retinal image. Up to now, psychophysical near-threshold measurements have been taken to suggest that observers select the maximum response from a bank of parallel linear visual filters, each sensitive to a specific image resolution, to perform detection. However, spatial-frequency tuned neurons in primary visual cortex are neither linear, nor independent and ample evidence emphasizes that perceptual decisions are mediated by pooling responses of multiple neurons. Why then does the aforementioned model do so well in explaining pattern detection? One possibility is that near-threshold stimuli are too weak to drive the early visual system's nonlinearities and activate only few sensory neurons. Alternatively, the ability of this theory to account for threshold experiments modelled in isolation belies the fact that its assumptions about pattern detection are inherently wrong. Here, we challenge both a linear channel model (LCM) and a neural population model (NPM) to fit a broad range of well-known and robust psychophysical pattern detection results, using a single set of parameters. In the LCM, psychophysical decisions reflect maximum-output decoding of linear and independent spatial frequency channels. In the NPM, perceptual choice behaviour is driven by maximum likelihood decoding of a population of normalized spatial-frequency tuned units resembling V1-neurons. We find that the LCM fails to satisfactorily explain pattern detection. The NPM, on the other hand, can fully account for pattern detectability as investigated in behavioural summation, adaptation and uncertainty experiments. This work thus offers a new theoretical interpretation for the vast psychophysical literature on pattern detection in which both normalization and maximum-likelihood decoding turn out to be crucial.