How Does Perceptual Discriminability Relate to Neuronal Receptive Fields?

Abstract

Perception is an outcome of neuronal computations. Our perception changes only when the underlying neuronal responses change. Because visual neurons preferentially respond to adjustments in some pixel values of an image more than others, our perception has greater sensitivity in detecting change to some pixel combinations more than others. Here, we examined how perceptual discriminability varies to arbitrary image perturbations assuming different models of neuronal responses. In particular, we investigated that under the assumption of different neuronal computations, how perceptual discriminability relates to neuronal receptive fields - the change in pixel combinations that invokes the largest increase in neuronal responses. We assumed that perceptual discriminability reflects the magnitude of change (the L2 norm) in neuronal responses, and the L2 norm assumption gained empirical support. We examined how perceptual discriminability relates to deterministic and stochastic neuronal computations. In the case of deterministic neuronal computations, perceptual discriminability is completely determined by neuronal receptive fields. For multiple layers of canonical linear-nonlinear (LN) computations in particular (which is a feed-forward neural network), neuronal receptive fields are linear transforms of the first-layer neurons' image filters. When one image is presented to the neural network, the first layer neurons' filters and the linear transform completely determine neuronal receptive fields across all layers, and perceptual discriminability to arbitrary distortions to the image. We expanded our analysis to examine stochastic neuronal computations, in which case perceptual discriminability can be summarized as the magnitude of change in stochastic neuronal responses, with the L2 norm being replaced by a Fisher-information computation. Using a practical lower bound on Fisher information, we showed that for stochastic neuronal computations, perceptual discriminability is completely determined by neuronal receptive fields, together with how responses co-variate across neurons.

Quantifying and predicting chromatic thresholds

“…A single reference color in the chromatic diagram (the xyY color space) can be perturbed along any direction in the two-dimensional xy plane, assuming that all such perturbations are unit-length vectors, and all perturbation to the reference forms a unit circle. Previous experimental evidence supported that perceptual discriminability reflect the extent of change in underlying neuronal responses, and in particular, the extent of change is consistent with a summary using the L2 norm (Poirson and Wandell 1990;Knoblauch and Maloney 1996;J. Zhou and Chun 2022).…”

Comparing neural models using their perceptual discriminability predictions

“…Computing Fisher information is generally non-trivial for arbitrary distributions, and experimentally, neural response distributions are rarely verified beyond the first two moments. In practice, a lower bound on Fisher information [19] can be used to summarize perceptual discriminability [20, 3]. This lower bound computation only involves the first two moments of the neural response distribution P ( r | s ).…”

“…Stimulus discriminability is the most reliable, widely-used, and well-understood measurement of perception. Discriminability is believed to reflect the degree of change in neural responses induced by small stimulus perturbations [1, 2, 3], and methods for estimating perceptual discrimination are highly refined and efficient. In this paper, we propose an efficient method to compare neural models based on their ability to account for human perceptual discriminability.…”

Delayed normalization model captures disparate nonlinear neural dynamics measured with different techniques in macaque and human V1

“…To understand the neuronal basis of visual perception, it is important to quantify the dynamics of neuronal responses and to characterize nonlinearities within these dynamics. This is because nonlinearities are central for our rich and flexible visual perception [8].…”