Research

研究活動

セミナー詳細

2010年02月24日

A three-dimensional spatio-temporal model of MT neurons that predicts responses to natural movies
Decoding visual experiences from brain activity evoked by natural movies

日 時 2010年02月24日(水) 13:30 より 15:00 まで
講演者 西本伸志先生
講演者所属 カリフォルニア大学バークレー校Jack Gallan研究室Postdoctoral fellow
お問い合わせ先 小松英彦(感覚認知情報研究部門 内線7861)
要旨

トピック1要旨
Area MT is an important site of motion processing, and many studies have used simple parametric stimuli to investigate mechanisms of motion processing in MT. How do area MT neurons encode information under natural viewing conditions? To address this issue we recorded from single MT neurons during stimulation with complex movies that simulate natural visual stimulation. We used these data to estimate spatio-temporal receptive field profiles for each neuron. Spectral profiles of individual neurons tend to lie on a single plane in the three-dimensional frequency domain, but the coverage of the plane differs across neurons. Most neurons form a partial ring in the plane, avoiding the region near zero temporal frequency. This suggests that MT neurons are optimized to represent velocity information in natural movies while simultaneously minimizing false alarms from static textures oriented parallel to the direction of motion.

トピック2要旨
Decoding perceptual experiences from brain activity is a challenging goal in applied neuroscience. Several recent studies have shown that it is possible to decode static monochromatic images from BOLD signals measured using fMRI. However, BOLD signals are extremely slow, so a decoder that can reconstruct continuous perceptual experiences from BOLD signals has been assumed beyond reach. Here we present a Bayesian decoder that combines voxel-based motion-energy encoding models with an implicit natural movie prior in order to reconstruct perceptual experiences from BOLD signals evoked by continuous natural movies. To construct this decoder we first evaluated several motion-energy encoding models. Each model was fit to BOLD signals of single voxels in posterior and ventral occipito-temporal visual cortex of fixating human subjects who viewed natural movies. We found that an encoding model that includes local directional motion information provides the most accurate predictions of BOLD signals. We then combined this motion encoding model with an implicit natural movie prior in order to reconstruct movies from BOLD signals. Our Bayesian decoder provides reconstructions of the spatio-temporal structure of natural movies. Further applications of the encoding and decoding framework will be discussed.