CRCNS: Online optimization for probing high-level auditory representations

生物学上重要的声音，如动物发声、语音和声调音乐，都包含丰富的和声 频谱能量聚集在基频的整数倍处。尽管准确的神经编码 谐音的产生机制尚不清楚，但最近的实验表明，谐音的敏感性是 广泛分布在绒猴的听觉皮质中。由于皮质谐波敏感度跨越多个八度， 它可能是通过组合典型地只偏爱单一频率的皮质下输入而得到的。我们 提出通过同时记录多个信号来研究猕猴听觉皮层的谐音编码 通过在线自适应刺激优化过程自动探测单个神经元 关于潜在神经回路的显式计算模型。传统的方法不能完全 表征由于刺激空间的组合爆炸而产生的复谐响应， 这是感官编码研究的普遍障碍。我们建议使用一种 自适应在线谐波刺激设计方法。我们将应用两种广泛的方法，一种是找到 一种是最好地驱动神经元的最佳刺激，另一种是基于模型的刺激设计，它可以有效地 确定每个给定的模型，并通过找到最能区分它们的刺激因素来比较竞争模型。我们 将开发：(1)自动化系统，以表征单个神经元在多个 使用神经像素记录探头的听觉皮质层，(2)用于表征谐波的自动化系统 使用双光子成像技术研究听觉皮质对多个八度频率的灵敏度，以及(3) 听觉皮质中有效编码谐音的产生电路模型。通过将刺激限制在 和声，这是足够复杂，但仍然容易处理，我们认为我们的方法更有可能 取得重大成功。我们已经在几个成功的网上取得了有希望的初步结果 神经生理学实验采用单单元记录法。将我们的在线方法扩展到神经像素 记录和双光子成像是合乎逻辑的下一步，可能会使许多正在工作的研究人员受益 在相关问题上。我们希望获得皮层神经元的完整刺激反应场景，以及 推断的电路模型，可以解释更高级别的谐波皮质表示如何 从更简单的输入组件产生，所有这些表示都将在高效的上下文中进行检查 对自然声音进行编码。

Biologically important sounds, such as animal vocalization, speech, and tonal music, contain rich harmonics with spectral energy clustered at integer multiples of the fundamental frequency. Although the exact neural coding mechanisms for harmonic sounds remain unclear, recent experiments show that harmonic sensitivity is widespread in the auditory cortices of the marmoset. Since cortical harmonic sensitivity spans multiple octaves, it is derived presumably by combining subcortical inputs that typically prefer only a single frequency. We propose to study harmonic coding in auditory cortex of the marmoset by simultaneous recording of many individual neurons which are probed automatically by an online adaptive stimulus optimization procedure based on explicit computational models of the underlying neural circuits. Conventional methods are incapable of fully characterizing complex harmonic responses because of the combinatorial explosion of the stimulus space, which is a general obstacle for sensory coding research. We propose to overcome this obstacle using an adaptive online approach to harmonic stimulus design. We will apply two broad types of methods, one is to find the optimal stimulus that best drive a neuron, and the other is model-based stimulus design that can effectively identify each given model and compare competing models by finding the stimuli that best distinguish them. We will develop: (1) automated system to characterize harmonic sensitivity of individual neurons across multiple layers of auditory cortex using Neuropixels recording probes, (2) automated system to characterize harmonic sensitivity in auditory cortex across multiple octaves of frequencies using two-photon imaging, and (3) generative circuit models for efficient coding of harmonic sounds in auditory cortex. By restricting the stimuli to harmonic sounds, which are complex enough but still tractable, we believe our methods are more likely to achieve significant success. We have obtained promising preliminary results in several successful online neurophysiological experiments using single-unit recording. Extending our online methods to Neuropixels recording and two-photon imaging is a logical next step and may potentially benefit many researchers working on related problems. We expect to obtain full stimulus-response landscapes of cortical neurons together with inferred circuit models that may explain how exactly a higher-level cortical representation of harmonics may arise from simpler input components, and all these representations will be examined in the context of efficient coding of natural sounds.