Abstract
Theoretical models of unsupervised category learning postulate that humans "invent" categories to accommodate new patterns, but tend to group stimuli into a small number of categories. This "Occam's razor" principle is motivated by normative rules of statistical inference. If categories influence perception, then one should find effects of category invention on simple perceptual estimation. In a series of experiments, we tested this prediction by asking participants to estimate the number of colored circles on a computer screen, with the number of circles drawn from a color-specific distribution. When the distributions associated with each color overlapped substantially, participants' estimates were biased toward values intermediate between the two means, indicating that subjects ignored the color of the circles and grouped different-colored stimuli into one perceptual category. These data suggest that humans favor simpler explanations of sensory inputs. In contrast, when the distributions associated with each color overlapped minimally, the bias was reduced (i.e., the estimates for each color were closer to the true means), indicating that sensory evidence for more complex explanations can override the simplicity bias. We present a rational analysis of our task, showing how these qualitative patterns can arise from Bayesian computations.
Introduction
In a complex world, how do we decide which stimuli belong together? While many studies look at how we learn pre-defined categories (like "dog" vs. "cat"), we often face "unsupervised" situations where we must discover the underlying structure of data ourselves. This study investigates whether the "simplicity principle"—the idea that humans favor the fewest possible explanations for data (Occam's razor)—influences basic perception. Specifically, the authors ask: if two different-colored sets of objects have similar statistical properties, will our brains merge them into a single category? And if they do, how does that merging distort our actual estimate of the objects' properties?
Methods
The researchers conducted three behavioral experiments where participants viewed a random scattering of colored circles on a screen for a brief moment. Participants had to estimate the number of circles without counting them.
The Manipulation
The number of circles on each trial was drawn from Gaussian distributions specific to the color of the circles. The authors varied the "overlap" between these distributions. In "High Overlap" conditions, the distributions for different colors were statistically similar. In "Low Overlap" conditions, they were distinct.
Computational Modeling
Human performance was compared against a "rational analysis" model based on Bayesian inference using a Chinese Restaurant Process (CRP) prior. This prior allows the model to "invent" an unbounded number of categories but penalizes the creation of new ones, favoring simplicity. This was compared to an exemplar model, which treats every trial as a unique category.
Key Results
Perceptual Regularization
When distributions overlap significantly, participants' estimates are "pulled" toward the average of the two distributions. This suggests the brain merges the two colors into one functional category to save on model complexity. In Experiment 1, for a baseline color with a mean of 65, estimates were significantly lower (~62.25) when paired with an alternative color mean of 55 (high overlap). When the alternative mean was 35 (low overlap), this bias was significantly reduced.
Evidence Overrides Simplicity
Occam’s razor is not absolute. When statistical evidence for separate categories becomes strong (i.e., distributions are far apart), the brain "splits" the categories, and the perceptual bias decreases. Experiment 2 further confirmed that increasing distribution variance (and thus overlap) increased the bias, reinforcing that statistical uncertainty promotes category merging.
Model Comparison
Quantitative model fitting showed that the rational CRP-based model significantly outperformed the exemplar model across all experiments. The fitted model preferred a small number of categories (median = 6) rather than treating every trial as unique, suggesting that human perception is the result of an active statistical inference process favoring simplicity.
Discussion
The findings suggest that humans do not just perceive raw data; we "invent" categories to simplify our world. This simplicity principle biases our perceptual estimates toward a "regularized" average. However, there are limitations to the current model. Behavioral data showed significant trial-to-trial correlations, which the primary rational model (assuming exchangeable trials) does not fully capture. Additionally, while the study focused on numerosity, it did not strictly control for other visual cues like density. Future work could explore how these simplicity biases are modulated by environmental complexity and the specific algorithmic heuristics humans use to approximate Bayesian inference.
Summary updated