Dialog concepts for learning retina encoders

This paper defines the biomedical and technological constraints for retina implants and describes our recent results and concepts regarding the function and training of learning retina encoder (RE) by means of neural networks. We show that: 1) system identification of a portable Mark I RE in analog hardware with tunable RF-filters can be achieved by means of supervised learning in connection with a neural net; and 2) the tunable RE-filters of a Mark II RE as DSP implementation can be tuned individually to typical RF properties of primate retinal ganglion cells in real time. These results demonstrate the feasibility of approximating the retinal information processing by means of tunable RF-filters in a RE system with a photosensor array at the input and simulated ganglion cells at the output. A concept is presented for training the learning RE with tunable RE-filters in a dialog with human subjects using their visual perception