5.4. Discussion 5 73 Figure 5.8: Results of training our simulator in an end-to-end pipeline on naturalistic images from the ADE20K dataset (Zhou et al., 2019). In the constrained optimization condition and the supervised boundary reconstruction condition, the encoder was configured to output 10 discrete stimulation amplitudes within the safe range of stimulation (0 to 128µA). The selected images represent the first three categories in the validation dataset (’Abbey’, ’Access Road’, ’Airbase’). Note that the brightness is enhanced in the phosphene images of the constrained optimization and the supervised boundary condition by 40%. (SeeFigureS6andFigureS7for enlarged and inverted SPV images) neuroprosthetics studies over the past decades. When we used a GPU, the simulator ran in real time, and as such, it could be used in experiments with sighted volunteers. Furthermore, our proof-of-principle computational experiments, presented in section Section 5.3.3, demonstrate the suitability of the simulator for machine learning pipelines, aimed at improving the image processing and stimulation strategies. Here we discuss some implications of our findings. 5.4.1. Validation experiments Visuotopic mapping The results presented inFigure5.1illustrate the value of including a visuotopic model based on spread of cortical activation to realistically estimate phosphene locations and size. Some previous studies have used a model of cortical magnification (Paraskevoudi & Pezaris, 2021; Srivastava et al., 2009) or visuotopic mapping (Fehervari et al., 2010; Li, 2013) in their phosphene simulations. However, our simulator is the first to incorporate empirical models of the current spread in cortical tissue (Bosking et al., 2017a; Tehovnik & Slocum, 2007; Winawer & Parvizi, 2016) to simulate the effects of stimulation current on the phosphene size. The accurate modelling of this biophysical relationship can help to increase the validity of simulation studies and brings fundamental SPV research closer to addressing questions regarding the practical real-life requirements of a visual prosthesis. Furthermore, the explicit link between the modeled area of cortical activation and the simulated phosphene sizes and locations makes our software suitable for including new receptive field modeling results (seeFigureS1for an example simulation based on 3D receptive field modelling using third party software). Future studies that target other structures than V1 that contain a retinotopic map, such as the LGN, can also use the simulator by replacing the V1 map with a retinotopic map of the respective brain structure. Collaborative international projects such as the PRIMatE Resource Exchange (PRIME-RE)

RkJQdWJsaXNoZXIy MTk4NDMw