Abstract
Abstract: :
Purpose: To develop and compare learning algorithms for a tunable retina encoder to simulate retinal information processing in learning retina implants. Methods: A retina encoder (RE) was implemented by an array of 256 tunable spatio-temporal filters (ST) with receptive field properties of retinal neurons to map visual patterns P1 onto encoded output patterns. RE-output signals were fed into a central visual system model (VM) to simulate the mapping of retinal output time courses onto visual percepts P2, which were visualized on a monitor. Alternative tuning algorithms (TA) were developed to roam within the large spatio-temporal parameter space of RE. Human volunteers with normal vision were asked to approximate P2 to P1 by means of TA-inputs via a set of manually controlled slide tracks, dials, or a keyboard. P1 was presented as visual pattern or was announced via the auditory channel. Results: (1) VM as neural network successfully learned to generate P2 with a small modified Hamming distance (range: 0-256) between P2 and P1 for various selected parameter vectors of RE, thus simulating the central visual system of various blind individuals. (2) Dialog-based RE tuning was successfully tested in subjects by presenting a given small set of P1 to the combination of a RE (initially tuned at random) and a VM (following learning) and by monitoring the gradual decrease of the Hamming distance down to values below 30, while subjects suggested TA changes so as to continuously approximate P2 to P1. (3) The generalization capability of the learning RE following RE tuning with a small set of input patterns P1 was successfully tested by demonstrating good quality output pattern time courses P2(t) in response to video camera inputs as P1(t). (4) RE tuning based on specific learning algorithms was compared with 'Random Search' and 'Manual Parameter Exploration'. Comparative data will be presented. Conclusions: A.Dialog-based RE tuning in learning retina implants can be tested in subjects with normal vision. B.Arrays of tunable spatio-temporal filters are powerful simulators for parts of the visual system. C.Pre-tuning before the application of learning algorithms can improve the final tuning quality and reduce the tuning time.
Keywords: pattern vision • computational modeling • retinal connections, networks, circuitry