# Generative modelling of continuous feature foraging reveals probabilistic representations of target distributions

## TL;DR

The study shows how people use probabilistic models to focus on important visual items in complex environments.

## Contribution

A novel foraging task and Bayesian generative model reveal how people track probabilistic distributions of visual targets.

## Key findings

- Participants preferentially selected colors with higher probability in the scene.
- The model captures how observers infer distribution properties through continuous foraging.
- Selection likelihood aligns with the underlying probability distribution of target colors.

## Abstract

To successfully orient ourselves within noisy visual environments, we must focus our attention on items of importance, ignoring sources of distraction. This selective attending is typically thought to be facilitated by templates, tuned towards current goals. However, in real-world scenes, the appearance of objects, such as their colour or luminance, varies greatly due to perceptual interpretation and environmental factors. Therefore, tuning attentional templates probabilistically may be more efficient than tuning them to precise values. This seems particularly important during continuous tasks, that require the selection of multiple objects which share certain properties. We investigated the effects of variability in target identity, using a novel foraging task. Participants (N = 15) had to continuously select 30 target objects, drawn from a truncated Gaussian colour distribution, sampled from a linearized space of 48 isoluminant hues. We adapted a generative model and applied it to the data, within a Bayesian multilevel framework. The model characterizes foraging as a sampling process without replacement and allows us to break foraging down into behavioural patterns that influence individual's target selection, independent of the number of targets present. The modelling results demonstrate increased likelihood of selection of more probable colour values in the scene. This likelihood maps onto the underlying probability distribution, illustrating how observers can acquire knowledge of the distribution's properties through foraging, beyond just the summary statistics.