5.3. Results 5 67 Figure 5.2: Estimate of the relative phosphene brightness for different stimulation amplitudes. The simulator was provided with a stimulation train of 166ms with a pulse width of 170µs at a frequency of 300Hz (see Equations(5.7) to (5.10)). Left: the fitted brightness levels reproduced by our model (red) and psychometric data (light blue) reported byFernández et al. (2021). Note that for stimulation amplitudes of 20.0µAand lower, the simulator generated no phosphenes as the threshold for activation was not reached. Right: the modeled tissue activation and brightness response over time. Figure5.4displays the simulator’s fit on the temporal dynamics found in a previous published study by (Schmidt et al., 1996). Here, cross-validation was not feasible due to the limited amount of quantative data. For repeated stimulation at different timescales (intervals of 4 seconds, and intervals of orange 200 seconds), the brightness of a single phosphene is evaluated after fitting the memory trace parameters. The observed accommodation effects in the simulator are compared to the data fromSchmidt et al. (1996).FigureS5shows the effect of the modeled temporal dynamics on the simulator’s output for continuous stimulation over 7 seconds. 5.3.2. Performance We tested the computational efficiency of our simulator, by converting a pre-processed example video2 (1504 frames) into simulated phosphene images, for different numbers of phosphenes, and at varying image resolutions. The simulator was run on a CUDAenabled graphics card (NVIDIA© A30) and each setting was run five times. The results are displayed inFigure5.5. The lowest measured frame rate (10.000 phosphenes at a resolution of 256×256) was 28.7 frames per second. Note that the missing combinations inFigure5.5indicate that the required memory exceeded the capacity of our GPU, as the simulation of large numbers of phosphenes at high resolutions can be memory intensive. Notably, even on a consumer-grade GPU (e.g., a 2016 model GeForce GTX© 1080) the simulator still reaches real-time processing speeds (>100 fps) for simulations with 1000 phosphenes at 256x256 resolution. 5.3.3. Usability in a deep learning SPV pipeline To validate that the simulator can conveniently be incorporated in a machine learning pipeline, we replicated an existing SPV pipeline by (de Ruyter van Steveninck et al., 2022a), replacing the simulator of that study with our biologically plausible simulator. We performed several phosphene encoding optimization experiments, described below. 2The example video with the simulated phosphene output can be downloaded via this link.
RkJQdWJsaXNoZXIy MTk4NDMw