3 Towards a task-based computational benchmark for the evaluation of prosthetic vision . Abstract In the near future, neuroprosthetic visual implants are expected to provide an opportunity to recreate an artificial form of visual perception in the blind. The developments are still ongoing and research is actively investigating different prototype designs and evaluating the expected functional outcomes. To augment brain stimulation research, many studies implement simulation paradigms to speed up the experimental cycle of hypothesis testing. However, there is a remaining need for fast and standardized evaluation benchmarks. Moreover, behavioral experiments can be costly and time-consuming and only a restricted number of study conditions can be compared. In this study, we propose a deep-reinforcement-learning-based computational framework that can serve as a complementary benchmark to behavioral simulation experiments for evaluating and optimizing prosthetic vision. The framework simulates a virtual implant user performing a mobility task in a 3D visual environment. Several experiments are performed that are explicitly modeled after a previously published simulation study with sighted human participants. We evaluated the effect of changing the edge detection threshold and the number of implanted electrodes as example cases of prototyping software and hardware parameters. The results illustrate that our computational framework evaluates these different implant characteristics in a similar way to the behavioral baseline study. The evaluation of taskrelevant information content in a confined benchmark can provide a useful early-stage indication of the functional visual quality. Although presented findings are basic examples, they indicate that the-based framework can provide a fast and inexpensive addition to existing simulation research and clinical experiments. This chapter is in preparation for submission as de Ruyter van Steveninck, J., Danen, S., Küçükog˘lu, B., Güçlü, U., van Wezel, R., & van Gerven, M. (2024). Deep reinforcement learning for evaluation and optimization of prosthetic vision 25
RkJQdWJsaXNoZXIy MTk4NDMw