Abstract
Purpose: :
To study the combined effects that a) selective sequential presentations of partial patterns and b) selective electrode subgroup stimulation have on 'Gestalt' perception in a retina implant.
Methods: :
A: Simulations were performed with a filter module (FM) of 10 x 10 spatio-temporal (ST) filters and an inverter module (IM) to mimic parts of the retina and the central visual system (Borbe, Eckmiller: ARVO, 2008).B: A partial pattern selection module (PM) at the FM-input subdivided P1 into up to 3 partial patterns (Pj) for sequential presentation during a presentation time (TP).C: A ganglion (G) cell selection module (GM) between the FM-output and the IM-input specified stimulation parameters for each electrode. Each of the 10 x 10 electrodes could stimulate up to 3 neighboring retinal G-cells. Accordingly, IM had 3 x 100 accessible input channels.D: A timing coordination module (TM) specified corresponding functions in PM, FM, and GM as a function of time.E: A dialog module (DM) managed the perception-based tuning using feedback from a human subject and it controlled PM, FM, and GM via TM.
Results: :
(1) Initially, the fixed mapping properties of IM assured that IM generated P2 sufficiently similar to corresponding P1 if the various parameters of PM, FM, GM, and TM had been set to a given reference parameter vector (PVref).(2) Subsequently, the reference parameter vector was randomly changed so that P2 became very different from P1.(3) During the perception-based dialog, subjects with normal vision received auditory information regarding P1 and observed P2 on a monitor. An evolutionary algorithm was used to tune the parameters of PM, FM, GM, and TM. Gradually, P2 changed and appeared similar to P1.(4) After completion of the tuning phase that used a small number of training patterns P1, new P1 could typically be transformed into P2 similar to P1.(5) The number of tuning iterations required depended on the learning algorithm used and the initially selected PVref. Tuning could usually be achieved within several hours.(6) How the visual percept quality depended on the spatial and temporal parameters of pattern presentation and of electrode stimulation could be analyzed.
Conclusions: :
A. The dialog-based tuning capability of a learning retina implant can be extended to include sequential presentation of partial patterns (with PM) and selective G-cell stimulation (with GM).B. Given the small number and large size of stimulation electrodes currently available, the proposed temporal distribution of percept-inducing stimulus patterns (with PM and TM) can improve the quality of ‘Gestalt’ perception re-gained in future intelligent visual prostheses.
Keywords: pattern vision • retinal connections, networks, circuitry • computational modeling