Shape Learning: Computational Changes in Chronically Studied Neural Populations

 描述（由申请人提供）：学习辨别新形状是人类和其他灵长类动物的基本视觉能力。它依赖于灵长类动物视觉皮层腹侧通路中形状计算的长期变化，特别是在IT（下颞叶皮层）的最后阶段。我们的目标是在单个神经元和神经回路的水平上研究这些变化，通过（i）分析连续识别的神经群体在长时间尺度（数周至数月）内的渐进形状计算变化，以及（ii）将这些变化与形状识别准确性和速度的改善相关联。我们将通过结合我们两个实验室开发的方法来实现这一目标。康纳实验室开发了神经形状计算的数学分析，基于遗传算法指导的大规模自适应刺激采样和刺激几何的多维参数化。Leopold实验室已经开发出使用微线束植入物对IT神经元群体进行长期电生理记录的方法，这些神经元通过其在100 s刺激中的签名响应模式连续识别。自适应采样可以利用微线束的采样时间的数量级增加，为腹侧通路皮质中数学上易处理的物体刺激的高通量测试提供了一种新的范式。基于我们以前对形状编码和形状处理动力学的研究，我们假设学习准确快速地区分新形状是基于通过对该形状的组成片段进行不同的组合计算的进展：（i）初始低准确度行为反映了存在于未训练状态中的形状片段信号的线性组合，（ii）在早期学习期间增加的准确性反映了经常性的网络非线性计算，为形状片段组合产生缓慢但明确的信号;（iii）后期学习过程中速度的提高反映了前馈非线性计算，从而产生准确、快速的性能。长期的微线记录将使我们能够跟踪数十个单个神经元的计算过程，并将计算变化与时间上的行为改善相关联。这将是第一次连续观察在长时间的视觉学习（数周至数月）中单个IT细胞的计算变化。无论具体的假设是否得到验证，这将提供迄今为止最直接的见解，了解IT电路级信息处理的具体变化如何与形状学习相关，这对我们理解符号和物体至关重要。

 DESCRIPTION (provided by applicant): Learning to discriminate new shapes is a fundamental visual ability for humans and other primates. It depends on long-term changes in shape computations in the ventral pathway of primate visual cortex, especially at final stages in IT (inferotemporal cortex). Our goal is to investigate these changes at the level of individual neurons and neural circuits, by (i) analyzing progressive shape computation changes in continuously identified neural populations across long timescales (weeks to months) and (ii) correlating these changes with improvements in shape discrimination accuracy and speed. We would achieve this goal by combining methodologies developed in our two laboratories. The Connor lab has developed mathematical analyses of neural shape computations, based on large-scale adaptive stimulus sampling guided by genetic algorithms and multi- dimensional parameterization of stimulus geometry. The Leopold lab has developed the use of microwire bundle implants for long-term electrophysiological recording from populations of IT neurons, continuously identified by their signature response patterns across 100s of stimuli. Adaptive sampling can leverage the order of magnitude increase in sampling time with microwire bundles, offering a new paradigm for high- throughput testing of mathematically tractable object stimuli in ventral pathway cortex. Based on our previous investigations of shape coding and shape processing dynamics, we hypothesize that learning to discriminate a new shape accurately and rapidly is based on a progression through distinct combinatorial computations operating on that shape's constituent fragments: (i) Initial low-accuracy behavior reflects linear combination of shape fragment signals, present in the untrained state, yielding only ambiguous information about complex shape configurations; (ii) Increasing accuracy during early learning reflects recurrent network nonlinear computations, yielding slow but unambiguous signals for shape fragment combinations; (iii) Increasing speed during late learning reflects feed-forward nonlinear computations, yielding accurate, fast performance. Chronic microwire recording will allow us to track this computational progression, for dozens of individual neurons, and correlate computational changes with behavioral improvements through time. This would be the first continuous observation of computational changes in individual IT cells during extended periods of visual learning (weeks to months). Whether or not the specific hypotheses are verified, this will provide the most direct insights to date into how specific changes in IT circuit-level information processing relate to shape learning, which is critical to our understanding of symbols and objects.