Dynamic of gain-modulated neural circuits

描述(申请人提供)：皮层神经元的反应是为其提供突触输入的其他神经元活动的函数。理论模型允许这样的函数是非线性的，但它所作用的输入活动的组合通常被认为是线性的。这导致所有输入对输出具有相同的质量影响。然而，实验表明，神经元的行为经常与这幅图背道而驰，这是出了名的。一个例子是顶后神经元中视网膜和眼睛位置信号之间的相互作用，其中视网膜外信息调节由视觉刺激触发的反应的幅度。这通常被描述为两个项之间的乘法相互作用，一个驱动神经元，另一个调节其增益。经典和非经典感受场之间的区别是直接影响和调制影响的另一个例子。众所周知，增益调制单元的电路非常适合执行某些类型的计算，但它们的动力学实际上是未知的。通过将理论模型和计算机模拟相结合，我们将探索神经元之间的乘法相互作用产生的神经电路具有比传统模型更丰富、更强大、更接近现实的动态特性。将构建各种新的网络模型，其中输入被分成两类：一类是驱动目标神经元的，另一类是调制被驱动反应的幅度的。初始模型将基于平均射速描述，并将与没有显式增益交互作用的传统模型进行比较。具有固定连通性(均匀、随机、中心环绕)的电路将首先被研究。初步结果揭示了这些模型中极其有趣和强大的动态：它们可能放大弱输入，但避免失控行为；它们可能被设置在两个不同的反应水平，起到开关的作用；它们可能自然地产生自我维持的活动模式，就像在工作记忆中一样；它们可能会产生活动的行波。这些行为将被编入目录。表现出类似动力学的尖峰神经元电路也将被开发出来。这将需要研究潜在的增益调制的生物物理机制。一直以来，增益调制尖峰网络将被用来构建感觉和运动系统的详细模型，这些模型已经在实验中得到了很好的表征。

DESCRIPTION (provided by applicant): The response of a cortical neuron is a function of the activity of other neurons that provide synaptic input to it. Theoretical models allow such function to be nonlinear, but the combination of input activities that it acts upon is typically considered linear. This results in all inputs having the same qualitative effect on the output. However, experiments reveal that neuronal behavior often deviates notoriously from this picture. One example is the interaction between retinal and eye-position signals in posterior parietal neurons, where the extra-retinal information modulates the amplitude of responses triggered by visual stimuli. This is often described as a multiplicative interaction between two terms, one that drives the neuron and another that regulates its gain. The distinction between classical and extra-classical receptive fields is another example of direct versus modulatory influences. Circuits of gain-modulated units are known to be ideally suited for performing certain kinds of computations, but their dynamics are virtually unknown. By combining theoretical models and computer simulations, we will explore the idea that multiplicative interactions between neurons give rise to neural circuits with dynamical properties that are richer, more powerful, and closer to reality than those of traditional models. A variety of new network models will be constructed in which inputs are segregated into two classes: ones that drive the target neuron and others that modulate the amplitude of the driven response. Initial models will be based on mean firing rate descriptions, and will be compared to traditional models without explicit gain interactions. Circuits with stereotyped connectivities (uniform, random, center-surround) will be studied first. Preliminary results reveal extremely interesting and robust dynamics in these models: they may amplify weak inputs but avoid runaway behavior; they may be set at two distinct response levels, acting as a switch; they may naturally give rise to patterns of self-sustained activity, as in working memory; they may produce traveling waves of activity. These behaviors will be catalogued. Circuits of spiking neurons exhibiting similar dynamics will also be developed. This will require investigating the biophysical mechanisms underlying gain modulation. All along, gain-modulated spiking networks will be used to construct detailed models of sensory and motor systems that have been well characterized experimentally.