Using reinforcement learning as an alternate method to test and automatically optimize pre-processing parameters for prosthetic vision
Researchers are close to developing cortical electrode implants through which blind people can get some sort of sight back. This sight comes in the form of phosphene vision which are basically bright light dots that can be triggered when parts of the visual cortex on the brain are arti cially stimulated with electrical currents. The resolution of this phosphene vi- sion is still rather limited. There have been proposed and tested many pre-processing techniques to best utilize this limited resolution but it can still be improved. Many of these techniques are tested in experiments with normal sighted human participants who get to see simulated phosphene vision on a screen or with a virtual reality headset. Such experiments are time-consuming and can be expensive. As alternative, we tried using reinforcement learning to conduct a similar experiment as one of those experiments where human participants had to do a way nding task with simulated phosphene vision with di erent phosphene resolutions. The results we got were promising showing many similarities to the experi- ment with human participants. However, they were not exactly the same. Furthermore, we also did an experiment where we let the reinforcement learning automatically optimize a parameter, the threshold for the Canny edge detection, of the pre-processing method we used. We found that in the condition that we changed the light intensity of the environment there was a signi cant di erence in the threshold that was learned by the model. For the other conditions no signi cant di erence was found. We then validated the performance of the learned threshold and found that for the lowest light intensity this vastly improved the reward the agent scored, but not for the brightest light intensity. When visually inspecting the output that came out of the optimal learned threshold for the low- est light intensity we also found that it may not be entirely optimal for humans. These results show that using reinforcement learning can result in useful things in improving phosphene vision, however it still has to be researched further to optimally make use of it.
Faculteit der Sociale Wetenschappen