Research Project 3 - Theory and computation of internal state dynamics

研究项目3 -内态动力学理论与计算 负责人：Surya Ganguli博士和Krishna Shenoy博士（与大卫Sussillo博士） 项目摘要 这个研究项目将发展理论原理和计算方法来阐明如何 外部输入与不同的内部神经状态动力学相互作用以驱动基本神经计算， 包括：（1）通过皮层放大微弱的感觉输入， 自发的背景活动;（2）贝叶斯集成的多感官输入，以计算内部 外部状态变量及其不确定性的估计;（3）离散状态的触发和维持 决定稳定行为的内部吸引子状态。此外，我们将开发通用和广泛适用的 计算工具，以支持同时全光读写多SLM技术，开发于 RP 1，并推广到RP 2和RP 4。这些工具将：（1）采用最先进的系统识别 从内部状态动态的神经数据网络模型中算法提取的方法;以及（2）采用 基于模型的控制理论方法来识别感兴趣的光遗传学刺激模式， 揭示了计算洞察力，以及使控制，皮层电路的动态。我们的理论 以及用于系统识别、洞察和控制的计算工具，都将推动 实验，并进而通过这些实验的结果迭代地细化，跨越所有RP。在Aim中 1我们将发展理论，解释外部输入和内部自发活动如何相互作用， 具有多种细胞类型的神经网络，可以对大脑皮层的感知灵敏度设置基本限制。 网络.我们将在一个紧密的理论-实验循环中反复测试这些理论， 系统：小鼠V1（RP 1）和猕猴V1（RP 2）。这项平行研究将使我们能够阐明 收敛和发散性质的基本计算的感觉放大在两个不同的 皮层网络的规模差异很大。我们还将探讨自发性和非自发性之间的相互作用 而诱发的活动会受到各种内部状态变化的影响，包括注意力、口渴感、饱腹感和顶部- 与RP 1和RP 2中完成的实验紧密合作，对V1进行向下控制。在目标2中，我们将开发 具有基本突触和细胞特性的神经回路如何将联合收割机内部状态与 外部输入以执行贝叶斯证据整合。我们将反复测试和完善这些理论， 理论驱动的实验，对小鼠V1、MEC、RSC和RP 4中的海马中的位置进行贝叶斯更新， 和贝叶斯更新的证据在联合记录猕猴V1和FEF在RP 2。在目标3中， 开发我们的通用计算工具，用于系统识别，洞察和控制， 我们将通过研究不同的内部状态动态来验证它们，包括鲁棒性和灵活性， RP 1中小鼠OFC和RP 4中小鼠RSC中吸引子转换的贝叶斯积分， RP 4中的小鼠RSC和RP 2中猕猴V1-FEF环中的贝叶斯证据整合。因此，总体而言，RP 3 通过RP 1、RP 2和RP 4以及DSC的紧密双向反馈回路发挥统一作用。

Research Project 3 - Theory and computation of internal state dynamics Leads: Surya Ganguli PhD and Krishna Shenoy PhD (with David Sussillo PhD) Project Summary This research project will develop both theoretical principles and computational methods for elucidating how external inputs interact with diverse internal neural state dynamics to drive fundamental neural computations, including: (1) the generation of accurate percepts through cortical amplification of weak sensory inputs amidst spontaneous background activity; (2) Bayesian integration of multisensory inputs to compute internal estimates of external state variables and their uncertainty; and (3) the triggering and maintenance of discrete internal attractor states dictating stable behaviors. Additionally, we will develop general and widely applicable computational tools to empower the simultaneous all-optical read-write multi-SLM technology, developed in RP1 and generalized to RP2 and RP4. These tools will: (1) employ state of the art systems identification methods to algorithmically extract from neural data network models of internal state dynamics; and (2) employ model based control theoretic methods to identify interesting optogenetic stimulation patterns that can both reveal computational insights into, as well as enable control of, the dynamics of cortical circuits. Our theories and computational tools for systems identification, insight and control, will both drive the design of experiments, and in-turn be iteratively refined by the outcomes of these experiments, across all RP’s. In Aim 1 we will develop theories for how the interplay of external inputs, and internal spontaneous activity in spiking neural networks with multiple cell-types, can set fundamental limits on the perceptual sensitivity of cortical networks. We will iteratively test these theories in a tight theory-experiment loop across 2 homologous sensory systems: mouse V1 in RP1 and macaque V1 in RP2. This parallel study will enable us to elucidate both convergent and divergent properties of the fundamental computation of sensory amplification in two different cortical networks that differ drastically in scale. We will also explore how the interplay between spontaneous and evoked activity is modified by diverse internal state changes, including attention, thirst, satiety, and top- down control of V1 in tight collaboration with experiments done in RP1 and RP2. In Aim 2 we will develop theories for how neural circuits with basic synaptic and cellular properties can combine internal states with external inputs to perform Bayesian integration of evidence. We will iteratively test and refine such theories in theory driven experiments on Bayesian updating of position in mouse V1, MEC, RSC and hippocampus in RP4, and Bayesian updating of evidence in joint recordings of macaque V1 and FEF in RP2. And in Aim 3 we will develop our generalized computational tools described above for systems identification, insight and control, and we will validate them by studying diverse internal state dynamics, including the robustness and flexibility of attractor transitions in mouse OFC in RP1 and mouse RSC in RP4, Bayesian integration of position in mouse RSC in RP4 and Bayesian integration of evidence in macaque V1-FEF loops in RP2. Thus overall, RP3 plays a unifying role through tight bi-directional feedback loops with RP1, RP2, and RP4 and the DSC.