Fusing Foveal Fixations Using Linear Retinal Transformations and Bayesian Experimental Design

Area of Science:

• Computational Neuroscience
• Computer Vision
• Machine Learning

Background:

• Human vision uses a high-resolution fovea and peripheral vision with decreasing resolution.
• Integrating information from multiple fixations is crucial for scene perception.
• Existing models may not fully capture the geometric transformation of retinal input.

Purpose of the Study:

• To develop a computational model for fusing visual information from multiple fixations.
• To represent the retinal transformation of a fixation as a linear downsampling process.
• To frame the problem of selecting the next fixation as a Bayesian experimental design task.

Main Methods:

• Explicitly representing retinal transformation as linear downsampling of a latent image.
• Utilizing factor analysis (FA) and mixtures of FA models for exact inference.
• Applying Bayesian experimental design with the expected information gain criterion for saccade planning.

Main Results:

• Demonstrated the effectiveness of the linear transformation model on Frey faces and MNIST datasets.
• Enabled exact inference for latent variables in factor analysis models.
• Successfully formulated and solved the problem of choosing the next fixation point.

Conclusions:

• The proposed linear downsampling model accurately represents retinal transformations.
• This approach facilitates efficient analysis of scene representations from visual fixations.
• The Bayesian experimental design framework offers a principled way to optimize visual search strategies.