An Active Inference Model of Covert and Overt Visual Attention

TL;DR

Addresses modeling covert and overt visual attention under active inference by dynamically modulating sensory precisions to minimize free-energy. The approach uses a 2D visual pipeline with a VAE-based exteroceptive generator and a diagonal sensory-precision matrix $\boldsymbol{\Pi}_s$ governed by covert attention, enabling endogenous/exogenous attention and saccadic actions via gradient flows from the free-energy $F$. It demonstrates, on the Posner cueing task and a simple target-focus task, that exogenous/valid cues speed responses and that the model exhibits IOR-like suppression, with bottom-up control yielding faster overt orienting than top-down control. The work provides a computational framework for perception, attention, and action in autonomously navigating robots and suggests pathways to extend to multi-target attention and learning.

Abstract

The ability to selectively attend to relevant stimuli while filtering out distractions is essential for agents that process complex, high-dimensional sensory input. This paper introduces a model of covert and overt visual attention through the framework of active inference, utilizing dynamic optimization of sensory precisions to minimize free-energy. The model determines visual sensory precisions based on both current environmental beliefs and sensory input, influencing attentional allocation in both covert and overt modalities. To test the effectiveness of the model, we analyze its behavior in the Posner cueing task and a simple target focus task using two-dimensional(2D) visual data. Reaction times are measured to investigate the interplay between exogenous and endogenous attention, as well as valid and invalid cueing. The results show that exogenous and valid cues generally lead to faster reaction times compared to endogenous and invalid cues. Furthermore, the model exhibits behavior similar to inhibition of return, where previously attended locations become suppressed after a specific cue-target onset asynchrony interval. Lastly, we investigate different aspects of overt attention and show that involuntary, reflexive saccades occur faster than intentional ones, but at the expense of adaptability.

Figures (7)

• Figure 1: At the core of the proposed model are the beliefs about the causes of sensory inputs. These beliefs and action signals are updated through attractor goals and error updates to minimize free-energy. The dedicated bottom-up attention module regulates attention through dynamic sensory precisions.
• Figure 2: The center of the RBF is (-0.25, 0.0), while the error appears at (-0.75, 0.0). The $u$-component of the RBF center is pushed toward the error with the update $\boldsymbol{\dot\mu_u} = -0.839$.
• Figure 3: Trial sequence of events. The model is first initialized for 10 steps, then a cue appears for 50 simulation steps. The cue is then removed for a variable interval, known as cue-target onset asynchrony (CTOA). After that the target appears until it is detected by the model or 1000 steps have passed.
• Figure 4: Reaction times and their averages as a function of target distance from focus point (CTOA = 100 for each trial)
• Figure 5: Covert attention center (dashed lines) and sphere position beliefs (solid lines) during valid trials, for both endogenous and exogenous cues. The horizontal line is the true target distance from center, and the vertical lines indicate trial events as in Fig. \ref{['fig:setup']}: the cue appears at step 10, disappears at step 60, target appears at step 160.