Data-Driven Reconstruction and Encoding of Sparse Stimuli across Convergent Sensory Layers from Downstream Neuronal Network Dynamics
Data-Driven Reconstruction and Encoding of Sparse Stimuli across Convergent Sensory Layers from Downstream Neuronal Network Dynamics
复制标题
DOI：
10.1137/21m1403114
复制
发表时间：
2021-01-01
影响因子：
2.1
通讯作者：
Xuan, Alex
中科院分区：
文献类型：
--
作者：
Barranca, Victor J.;Hu, Yolanda;Xuan, Alex
神经元网络在大小、连接结构和大脑下游层的功能方面差异很大。这就提出了一个问题：当信息沿着沿着压缩和扩展路径重新编码时，信息是否会丢失。在这项工作中，我们开发了一种潜在的数据驱动机制，用于保存下游层神经元网络活动中的信息，该机制利用单个神经元对足够强的斜坡人工输入的反应的广泛线性，以适应整个网络的线性输入-输出映射。我们分析了几个家庭的两层神经元网络模型的动力学，其中的输入组件远远超过下游的神经元，如在压缩路径，并应用拟合映射结合压缩感知理论，以重建稀疏结构的刺激。输入-输出映射促进刺激重建，其仅使用响应于短时间持续时间内的输入的下游神经元放电率的测量，提供刺激恢复，即使当理论分析是棘手的或动力系统的控制方程如在实验中是未知的时。类似地，在不同的单神经元模型、网络耦合函数和图像类中获得准确的刺激重建。当均匀随机前馈连接被空间局部前馈连接类似于感受野取代时，得到了改进的重建。我们希望类似的原理可以在假肢和大规模网络连接重建中进行实验。
Neuronal networks vary dramatically in size, connectivity structure, and functionality across downstream layers of the brain. This raises the question of whether information is lost as it is re-encoded along compressive and expansive pathways. In this work, we develop a potential data-driven mechanism for the preservation of information in the activity of neuronal networks across downstream layers, which uses the widespread linearity of individual neuronal responses to sufficiently strong ramped artificial inputs to fit a linear input-output mapping across the network. We analyze the dynamics of several families of two-layer neuronal network models, where the input components far outnumber the downstream neurons, as in compressive pathways, and apply the fitted mapping in conjunction with compressive sensing theory to reconstruct stimuli with sparse structure. The input-output mapping facilitates stimulus reconstructions that only use measurements of downstream neuronal firing rates in response to inputs over a short time duration, furnishing stimulus recovery even when theoretical analysis is intractable or the governing equations of the dynamical system are unknown as in experiment. Similarly accurate stimulus reconstructions are obtained across different single-neuron models, network coupling functions, and image classes. Improved reconstructions are yielded when uniformly random feedforward connectivity is replaced by spatially localized feedforward connectivity akin to receptive fields. We expect that similar principles could be leveraged experimentally in prosthetics as well as in the reconstruction of large-scale network connectivity.