fMRI and Behavioral Studies of Unsupervised Learning in High Level Visual Cortex

描述（由申请人提供）：经验被认为在形成皮层表征方面起着关键作用，通过创造神经反应来支持物体识别，这些神经反应在某些变化维度上是选择性的，而在其他维度上是不变的。尽管许多先前的研究已经检查了监督训练对大脑目标选择区域的影响，但对于视网膜输入的统计规律在多大程度上可以直接塑造涉及目标识别的神经基质，我们所知甚少。无监督学习很重要，因为它允许大脑采用简单的自组织机制，将连续不断的视觉输入转化为我们经验的稳定对象。虽然行为和计算研究强烈表明无监督学习在物体识别中起着关键作用，但大多数研究输入统计作用的相关神经科学研究都集中在其对早期视觉区域的影响上。在这里，我们提出了结合功能磁共振成像、心理物理学和计算建模的前沿技术的实验，以检验关于物体识别中无监督学习的两个假设。首先，我们提出神经反应可能会被调整以匹配在无监督训练中经历的形状和对象样本的范围和频率。也就是说，在无监督训练中，相对于不经常看到或未训练的项目，神经反应会增加，并且对更频繁看到的项目更具选择性。这可能提供了一种机制，可以提高对最常见刺激的辨别性能。其次，行为和计算证据提出了一个有趣的假设，即大脑使用时空相关性作为一种手段，将属于同一物体的不同图像绑定在一起，允许在戏剧性的转换中识别同一物体，例如由于旋转而导致的外观变化。我们将确定在无监督训练期间，视觉输入的时空相关性是否会增加大脑反应和感知表现的不变性，这些反应和感知表现相对于以不相关方式训练的类似项目和训练前的反应（和表现）。第三，我们将检验非监督学习的机制是否可以推广到监督学习。在我们所有的实验中，我们将检查无监督训练前后的神经反应和表现，并使用计算建模将fMRI数据与可能的潜在神经机制（如神经调谐的锐化和放电率的增加）联系起来。这项工作将填补重要的知识空白，提供了从视觉经验统计中产生有效对象识别表征的神经机制的第一个说明。

Description (provided by applicant): Experience is thought to play a critical role in shaping the cortical representations that support object recognition by creating neural responses are selective for some dimensions of change and invariant to others. Although many previous studies have examined the effects of supervised training on object selective regions of the brain, much less is known about the degree to which statistical regularities in the retinal input can directly shape the neural substrates involved in object recognition. Unsupervised learning is important because it allows the brain to employ simple self organizing mechanisms that turn the continuous flux of visual input into the stable objects of our experience. While behavioral and computational work strongly suggests that unsupervised learning plays a key role in object recognition, most related neuroscience work examining the role of input statistics has focused on its effects in early visual areas. Here we propose experiments that combine cutting edge techniques in fMRI, psychophysics, and computational modeling to examine two hypotheses concerning unsupervised learning in object recognition. First, we propose that neural responses may become tuned to match the range and frequency of shape and object exemplars experienced during unsupervised training. That is, neural responses will increase and become more selective for items seen more frequently during unsupervised training relative to infrequently seen or untrained items. This may provide a mechanism which improves discrimination performance for stimuli seen most frequently. Second, behavioral and computational evidence suggests the intriguing hypothesis that the brain uses spatio-temporal correlations as a means for binding different images as belonging to the same object, allowing for recognition of the same object across dramatic transformations, such as changes in its appearance due to rotation. We will determine if spatio- temporal correlations in the visual input during unsupervised training increases the invariance of both brain responses and perceptual performance relative to similar items trained in an uncorrelated manner and pre- training responses (and performance). Third, we will examine if mechanisms of unsupervised learning generalize to supervised learning. In all of our experiments we will examine neural responses and performance both before and after unsupervised training, and use computational modeling to link fMRI data to the possible underlying neural mechanisms such as sharpening of neural tuning and increased firing rates. The proposed work will fill important gaps in knowledge by providing the first account of the neural mechanisms that generate effective representations for object recognition from the statistics of visual experience.