Machine Learning Techniques for Simulating Human Psychophysical Testing of Low-Resolution Phosphene Face Images in...

Area of Science:

• Biomedical Engineering
• Computer Science
• Neuroscience

Background:

• Human psychophysical experiments are crucial but time-consuming for evaluating artificial visual percepts.
• Modifications in hardware/software require repeated, labor-intensive testing.
• Machine learning (ML) offers a potential solution to streamline this evaluation process.

Purpose of the Study:

• To investigate the capability of standard ML models to replicate human performance in match-to-sample tasks using phosphene image stimuli.
• To assess ML model accuracy in predicting human facial recognition abilities based on low-resolution phosphene inputs.
• To explore ML's potential in reducing the need for extensive psychophysical testing in visual prosthetic development.

Main Methods:

• Developed ML models trained on a dataset of 3600 phosphene images of human faces.
• Compared ML model performance against human subject data from a psychophysical test involving 720 phosphene images and 36 participants.
• Analyzed ML model ability to predict human recognition performance for novel phosphene images.

Main Results:

• A superior ML model accurately mirrored human behavioral trends, predicting 8 out of 9 phosphene quality levels.
• ML models demonstrated the capacity to predict human recognition performance for untested phosphene images.
• The ML approach significantly streamlined the evaluation process compared to traditional methods.

Conclusions:

• Machine learning models can effectively replicate and predict human performance in artificial vision tasks using phosphene stimuli.
• This ML-driven approach has the potential to significantly accelerate the research and development of visual prosthetics.
• The findings highlight a paradigm shift towards more efficient evaluation methods in the field of artificial vision.