Active Fovea-Based Vision Through Computationally-Effective Model-Based Prediction.

Area of Science:

• Computational Neuroscience
• Computer Vision
• Machine Learning

Background:

• Visuomotor control involves selecting actions (saccades) to interpret visual scenes, often modeled as partially-observed environments.
• Existing action selection metrics like Salience, Infomax, and Variational Free Energy can be unified under an information-theoretic framework.

Purpose of the Study:

• To propose novel, computationally simpler action selection metrics for visuomotor control.
• To evaluate these metrics in a fovea-based visual scene decoding task.
• To demonstrate the efficacy of combining predictive control with deep learning predictors.

Main Methods:

• Modeling the visual scene as a partially-observed environment with a generative model.
• Developing Information Gain Lower Bound (IGLB) and Information Gain Upper Bound (IGUB) metrics.
• Implementing a fovea-based visual scene decoding setup with numerical experiments.

Main Results:

• Achieving state-of-the-art recognition rates with fovea-based saccadic exploration using minimal image data (<10%).
• Highlighting the sub-optimality of classical action-selection metrics when using coarse or faulty inference models.
• Developing a computationally effective predictive model using the IGLB objective, demonstrating competitive accuracy and robustness.

Conclusions:

• Fovea-based saccadic exploration combined with deep learning predictors offers significant advantages in visual scene interpretation.
• Novel metrics like IGLB and IGUB provide efficient alternatives for action selection in visuomotor tasks.
• The proposed methods demonstrate robustness and accuracy, even with simplified predictive models.