Machine Learning Method for Functional Assessment of Retinal Models.

Area of Science:

• Biomedical Engineering
• Computational Neuroscience
• Machine Learning

Background:

• Retinal prostheses aim to restore vision for blind individuals by simulating Retinal Ganglion Cell (RGC) responses.
• Accurate retinal models are crucial for developing effective visual prostheses.

Purpose of the Study:

• Introduce Functional Assessment (FA) to evaluate retinal models on visual understanding tasks.
• Assess the performance of machine learning classifiers fed with RGC responses from retinal models.

Main Methods:

• Developed a machine learning pipeline using traditional classifiers on RGC responses.
• Tested models on object and digit recognition tasks (CIFAR-10, MNIST, Fashion MNIST, Imagenette).
• Investigated the impact of task type, input data manipulation, and dataset structure on FA performance.

Main Results:

• Retinal model performance varied significantly across datasets (MNIST/Fashion MNIST >80% accuracy, CIFAR-10/Imagenette ~40%).
• Image splitting did not substantially improve model accuracy.
• Models with higher accuracy in standard RGC response prediction also performed better in FA.

Conclusions:

• FA provides a direct measure of visual perception quality for retinal models.
• Dataset structure significantly influences model performance in FA.
• FA is a valuable tool for comparing the effectiveness of different retinal models for vision restoration.