Ensemble averaging: What can we learn from skewed feature distributions?

Area of Science:

• Cognitive Neuroscience
• Computational Vision
• Psychophysics

Background:

• Observers accurately estimate ensemble features, but the underlying visual system mechanisms are debated.
• Existing models propose arithmetic averaging or "robust averaging" that downweights outliers.
• Previous work suggested robust averaging based on motion direction biases in skewed distributions.

Purpose of the Study:

• To investigate ensemble perception mechanisms for orientation averaging.
• To determine if skew-related biases in orientation averaging depend on the entire feature distribution's shape.
• To test a biologically plausible model of robust averaging and distribution-dependent biases.

Main Methods:

• Replicated systematic skew-related biases in orientation averaging across four experiments.
• Analyzed the influence of mean, mode, and overall distribution shape on bias magnitude.
• Tested a population coding model based on spatial pooling and neural receptive fields.

Main Results:

• Orientation averaging exhibited systematic biases related to feature distribution skew.
• Bias magnitude was significantly influenced by the entire distribution's shape, not just mean or mode.
• Both a loss function model and a population coding model accurately predicted the observed biases.

Conclusions:

• Ensemble perception of orientation involves robust averaging influenced by distribution shape.
• A population coding model with winner-take-all decoding provides a neural implementation for these processes.
• Findings advance understanding of how the visual system samples and averages information from object ensembles.