Related Experiment Videos
Diagnostic features predicted by deep learning improve human object recognition in simulated prosthetic vision
Elsa Scialom1, Ben Lonnqvist2, Alban Bornet3
1EPFL, EPFL SV BMI LPSY SV 2805 (Bâtiment SV) Station 19, Lausanne, VD, 1015, Switzerland.
Journal of Neural Engineering
|July 29, 2026
Summary
Deep learning optimizes visual prostheses by identifying key features for object recognition, reducing the needed phosphenes for better prosthetic vision, especially with limited electrodes.
Area of Science:
- Neuroscience
- Computer Science
- Biomedical Engineering
Background:
- Cortical visual prostheses aim to restore vision via V1 electrical stimulation.
- Effectiveness is limited by the number of simultaneously elicitable phosphenes.
- Identifying sparse, informative features is crucial for functional prosthetic vision.
Purpose of the Study:
- To target diagnostic features for object recognition in visual prostheses.
- To optimize phosphene placement using a deep learning pipeline.
- To enhance object recognition performance under simulated prosthetic vision.
Main Methods:
- A deep learning pipeline was combined with bioplausible phosphene simulation.
- Phosphene renderings were optimized for object recognition.
- 63 participants performed object recognition tasks with simulated prosthetic vision.
Main Results:
- The DNN-based phosphene placement significantly reduced the number of phosphenes needed for object recognition.
- This advantage was most pronounced at low electrode densities.
- DNN approach outperformed contour-based and luminance-based placements.
Conclusions:
- Deep neural networks can identify diagnostic object features for enhanced recognition in visual prostheses.
- The proposed pipeline offers a principled method for optimizing visual representations.
- Task-relevant feature selection is vital given stimulation capacity constraints.