主催: The Japanese Society for Artificial Intelligence
会議名: 2024年度人工知能学会全国大会(第38回)
回次: 38
開催地: アクトシティ浜松+オンライン
開催日: 2024/05/28 - 2024/05/31
Perceptual illusions, like the rubber-hand and ventriloquism, demonstrate that simultaneous but spatially separated multisensory stimuli are combined into a single unified percept, but only if they are not too far apart. Here we show that when a feedforward plastic network is trained on spatially congruent multisensory stimuli, it develops spatial tuning which facilitates the accurate reproduction of multiple empirical findings, providing the network endogenously generates activity that is just above a transition from a stable fixed point state to a chaotically fluctuating one. Training selectively enhances the activity of some multisensory neurons and silences others, suppressing chaos specifically for those congruent multisensory stimulus combinations. After training, multisensory stimuli that fall in a distinct low spatial disparity regime activate highly reproducible fixed point attractor responses associated with a unified percept, while the network states activated by more spatially disparate multisensory stimuli fluctuate inconsistently across trials and represent segregated percepts.