Adaptive Spatial Calibration and Co-Adaptation Under Binary Feedback for Sensory Substitution

Abstract

Sensory substitution aims to convey visual information through another sense. Proposed corneal interfaces, for example, would translate visual patterns into spatial tactile cues. Keeping these cues aligned with their intended locations may require recalibration as the device response and the user's interpretation change. We investigate this problem in simulation using occasional left/right and above/below feedback. Six calibration methods are tuned on separate simulations and compared across 126 conditions, with 1,000 paired simulated users per condition. The main user-learning model receives the same noisy directional reports as the device. Across 20 core conditions, a dynamic logistic filter that tracks uncertainty reduced mean spatial error from 0.1131 to 0.1058 compared with fixed-size updates, an average reduction of 6.5%. It also increased the fraction of trials within an illustrative target radius of 0.10 from 50.6% to 54.7%. Its advantage persisted across the tested response-model mismatches, but reversed in four conditions with doubled drift. Allowing the simulated user and device to learn from the same noisy reports sometimes increased error. These findings show why calibration methods should be evaluated together with the feedback available to each learner and the speed of change. An expanded search over 111 candidates at 120 and 480 trials selected weak forgetting and slower decay for longer sessions, while faster drift still reversed some rankings. Together, the experiments provide a computational benchmark for selecting calibration policies in sensory-substitution research.