No abstract

Figure 1: A) Simulated prosthetic vision (SPV). Constrained by neuroanatomical and/or psychophysical data, a phosphene model Ψ (e.g., [2]) predicts what a retinal implant user should "see" for any given input stimulus. The predicted percept is typically a nonlinear continuous function of the input stimulus, thus Ψ can be approximated by a generic feedforward neural network (FNN), Ψ, which is amenable to differentiation. B) End-to-end optimization of bionic vision. For a given target percept, a stimulus encoder based on a convolutional neural network (CNN) is trained to predict the combination of active electrodes that generates the percept with the smallest possible reconstruction loss. The FNN approximator is fixed during encoder training.