Complexity and specificity of experimentally-induced expectations in motion perception

Nikos Gekas, Matthew Chalk, A. R. Seitz, Peggy SerièsJournal of Vision 2013 · DOI

Abstract

Our perceptions are fundamentally altered by our expectations, i.e., priors about the world. In previous statistical learning experiments, we investigated how such priors are formed by presenting subjects with white low contrast moving dots on a blank screen and using a bimodal distribution of motion directions such that two directions were more frequently presented than the others. We found that human observers quickly and automatically developed expectations for the most frequently presented directions of motion. Here, we examine the specificity of these expectations. Can one learn simultaneously to expect different motion directions for dots of different colors? We interleaved moving dot displays of two different colors, either red or green, with different motion direction distributions. When one distribution was bimodal while the other was uniform, we found that subjects learned a single bimodal prior for the two stimuli. On the contrary, when both distributions were similarly structured, we found evidence for the formation of two distinct priors, which significantly influenced the subjects' behavior when no stimulus was presented. Our results can be modeled using a Bayesian framework and discussed in terms of a suboptimality of the statistical learning process under some conditions.

Introduction

The "Bayesian Brain" hypothesis suggests that our perception is not a direct recording of sensory input but a combination of noisy sensory data and internal expectations (priors). While previous research showed that humans can quickly learn simple priors—such as the frequency of specific motion directions—real-world environments are complex and multidimensional. This study addresses the question of specificity: Can the brain simultaneously learn feature-contingent priors, such as expecting different motion directions for different colors? Understanding this reveals the constraints of our statistical learning mechanisms and how different processing streams in the brain, like color (ventral) and motion (dorsal), interact.

Methods

The researchers conducted two psychophysics experiments using low-contrast moving dot stimuli. Participants performed two primary tasks: an estimation task, where they indicated the perceived direction of motion, and a detection task, where they reported the presence and color of the stimuli.

Experimental setup and stimulus distributions. (a) The trial sequence includes fixation, a 3000ms estimation phase, and a detection phase. (b) Probability distributions for motion direction, including uniform and bimodal structures.

In Experiment 1, one color followed a uniform distribution while the other followed a bimodal distribution. In Experiment 2, both colors had distinct structures (one bimodal, one trimodal), though their combined distribution was uniform. Behavioral data from naive participants (22 in Exp 1, 23 in Exp 2) were analyzed using Bayesian computational models to determine if subjects employed a single merged prior or separate, color-specific priors.

Key Results and Insights

The Merging Tendency

In Experiment 1, subjects failed to maintain separate priors. Even when viewing the color associated with a uniform distribution, their perceptions were biased toward the bimodal distribution of the other color. This suggests a default "suboptimal" strategy where the brain merges expectations across features, potentially due to the computational cost of integrating information across separate cortical processing streams.

Structure Forces Specificity

In Experiment 2, when both colors possessed distinct statistical structures, participants were more successful at forming separate priors. This was evidenced by "hallucinations" during zero-stimulus trials, which were heavily clustered around the frequent directions specific to the perceived color.

Implicit Learning

Participants were generally unaware of the underlying statistics. Despite this lack of conscious awareness, their estimation biases and hallucinations (seeing motion when none existed) clearly reflected the learned distributions, highlighting the automatic nature of this statistical learning process.

Discussion

This research demonstrates that while the brain is capable of learning specific, feature-contingent expectations, it often defaults to a merged representation unless the statistical evidence for separation is strong. The difficulty in separating priors in Experiment 1 may be due to the relatively short training duration or the fact that color and motion are processed in different visual streams. Future work could investigate if learning is more efficient when features are processed within the same stream, such as motion direction and speed. Identifying the neural circuits responsible for this merging of priors remains a key challenge in understanding the Bayesian brain.

Summary updated

Abstract

Our perceptions are fundamentally altered by our expectations, i.e., priors about the world. In previous statistical learning experiments (Chalk, Seitz, & Seriès, 2010), we investigated how such priors are formed by presenting subjects with white low contrast moving dots on a blank screen and using a bimodal distribution of motion directions such that two directions were more frequently presented than the others. We found that human observers quickly and automatically developed expectations for the most frequently presented directions of motion. Here, we examine the specificity of these expectations. Can one learn simultaneously to expect different motion directions for dots of different colors? We interleaved moving dot displays of two different colors, either red or green, with different motion direction distributions. When one distribution was bimodal while the other was uniform, we found that subjects learned a single bimodal prior for the two stimuli. On the contrary, when both distributions were similarly structured, we found evidence for the formation of two distinct priors, which significantly influenced the subjects' behavior when no stimulus was presented. Our results can be modeled using a Bayesian framework and discussed in terms of a suboptimality of the statistical learning process under some conditions.
BibTeX
@article{gekas2013complexity,
  author = {Nikos Gekas and Matthew Chalk and A. R. Seitz and Peggy Seriès},
  title = {{Complexity and specificity of experimentally-induced expectations in motion perception}},
  journal = {Journal of Vision},
  year = {2013},
  doi = {10.1167/13.4.8},
}
Figure