Stojanoski Bobby, Cusack Rhodri
Brain and Mind Institute, University of Western Ontario, London ON, Canada Department of Psychology, University of Western Ontario, London ON, Canada.
Brain and Mind Institute, University of Western Ontario, London ON, Canada Department of Psychology, University of Western Ontario, London ON, Canada Medical Biophysics, University of Western Ontario, London ON, Canada.
J Vis. 2014 Oct 9;14(12):6. doi: 10.1167/14.12.6.
To isolate the neural mechanisms associated with recognizing objects from those processing basic visual properties, control stimuli are required that contain the same perceptual properties as the objects but are unrecognizable. We demonstrate that conventional methods for generating control stimuli (phase scrambling, box scrambling, texture scrambling) yield poor controls because they dramatically distort the basic visual properties (e.g., spatial frequency, perceptual organization) to which even the earliest stages of visual processing are sensitive. We developed a new scrambling method, using a diffeomorphic transformation that preserves the basic perceptual properties of the image while removing meaning. We acquired perceptual ratings to determine the least amount of scrambling necessary to remove recognition. We hypothesized that our "diffeomorphic" images would produce neural activity at the earliest stages of the visual system that more closely matched activity in response to intact images relative to the other scrambling methods. To test this hypothesis, we used the HMAX computational model of object recognition and compared the simulated neural activity at the earliest stages of the visual system (layers S1, C1, and S2) between a set of 149 images scrambled using each distortion method to their intact version. We found that scrambled "diffeomorphed" images were indistinguishable to intact images in each layer of the model, but all of the other distortion methods yielded quite different patterns. Our results indicate that "diffeomorphed" images serve as more appropriate control stimuli in neuroimaging studies that aim to disentangle the representations of perceptual and semantic object properties.
为了将与识别物体相关的神经机制与处理基本视觉属性的神经机制区分开来，需要控制刺激，这些刺激要包含与物体相同的感知属性，但无法被识别。我们证明，用于生成控制刺激的传统方法（相位扰乱、方块扰乱、纹理扰乱）产生的控制效果不佳，因为它们会极大地扭曲基本视觉属性（例如空间频率、感知组织），而即使是视觉处理的最早阶段也对这些属性敏感。我们开发了一种新的扰乱方法，使用一种微分同胚变换，该变换在去除意义的同时保留图像的基本感知属性。我们获取了感知评分，以确定消除识别所需的最少扰乱量。我们假设，相对于其他扰乱方法，我们的“微分同胚”图像在视觉系统的最早阶段产生的神经活动，会更紧密地匹配对完整图像的反应活动。为了验证这一假设，我们使用了物体识别的HMAX计算模型，并比较了在视觉系统最早阶段（S1层、C1层和S2层），149幅分别使用每种失真方法扰乱后的图像与其完整版本之间模拟的神经活动。我们发现，在模型的每一层中，扰乱后的“微分同胚”图像与完整图像无法区分，但所有其他失真方法产生的模式则大不相同。我们的结果表明，在旨在区分感知和语义物体属性表征的神经成像研究中，“微分同胚”图像是更合适的控制刺激。
