3 36 3. Towards a task-based computational evaluation benchmark et al. (2022b). Another notable correspondence is the drop off of the performance gain for higher phosphene resolutions. This corresponding finding suggests that for increasing numbers of electrodes the task-relevant visual information gain is inherently limited. The suggestion that adequate mobility performance can be achieved with a limited number of electrodes resonates with findings from related SPV research (Cha et al., 1992b; Srivastava et al., 2009). Hence, the implanted number of electrodes should be evaluated as a balanced trade-off between performance gain and other factors such as surgical risks. Environmental complexity The effects of environmental complexity in the current study were largely similar to the results in the baseline study, but there were also notable differences. In both studies, the plain environment yielded higher performance with low numbers of phosphenes. However, for higher numbers of phosphenes, the our results reveal no effect of visual complexity, whereas the baseline study found a significantly improved navigation speed in the complex environment. These differences are difficult to interpret, since we did not use navigation speed as output in our RL approach. Potential differences may result from inherent differences in information processing between the RL-based virtual implant users and the sighted subjects, or from differences in the task. These limitations are further discussed inSection 3.4.2. 3.4.2. General considerations Functonal hardware and software prototyping The functional benefits of visual prostheses depend on many choices in the hardware and software design. Clinical experiments are restricted by the limited quality of contemporary prostheses, and often test only basic visual function (e.g., seeChen et al., 2020; Fernández et al., 2021; Ho et al., 2015; Stingl et al., 2015). Task-based simulation experiments allow for functional evaluation of future implant designs in more naturalistic tasks. The example experiments with multiple phosphene resolutions and edge detection thresholds illustrate that our computational benchmark can be used to evaluate the effects of hardware and software parameters in complex naturalistic activities. As illustrated by the perturbation analysis, the virtual nature of the agent and environment allow for specific controlled experimentation. Note that the presented experiments are basic example cases, but the framework principally enables virtually unlimited exploration with different tasks, different image processing strategies, and different phosphene simulations. Virtual prosthesis users versus human subjects A critical feature as well as limitation is the use of virtual implant users that evaluate the quality of prosthetic vision. It is likely that the approach of sighted human subjects and blind implant users differs fundamentally from that of a trained RL agent. Convolutional neural networks have a different information processing architecture compared to the human visual system and they possess contextual knowledge about the world. Moreover, the virtual prosthesis users in our RL-based framework undergo a training procedure that ensures optimal interpretation of the information in the phosphene patterns. The training duration of these virtual agents (several hours) is in sharp contrast with the long
RkJQdWJsaXNoZXIy MTk4NDMw