Reconstructing neuro-dynamical principles of prefrontal cortical computations across cognitive tasks and species

本联盟提供了一个独特且史无前例的机会，以揭示跨物种、发育阶段和认知域的高级认知功能的神经动力学和计算原理和机制。该项目将利用这一机会，并将试图通过从统一动力系统的角度并在共同的统计机器学习框架内分析其他项目伙伴获得的实验数据来揭示一般计算原理。我们将使用深度生成递归神经网络(RNN)来逼近实验数据背后的动力系统。虽然RNN最近已成为神经科学中研究动力学机制的流行工具，但我们的方法更进一步，以统计、最大似然或贝叶斯方法直接从观察到的生理和行为时间序列推断RNN参数，产生定量预测。我们将与实验伙伴合作，利用我们小组建立的这些方法来研究决策的神经吸引子动力学，间隔计时背后的神经动力学机制，工作记忆的物种特有的神经计算机制，以及这些机制如何随着发育阶段的变化而变化。我们还将解决一个长期存在的问题，即在不同的认知任务中，类似或不同的神经计算原理是否构成依赖PFC的表现的基础。在每一种情况下，我们都提出了一组关于潜在神经动力学机制的具体假设，可以使用我们的方法进行测试。此外，我们还将生成具体的(定量)新颖预测，并反馈给实验合作伙伴。此外，我们的RNN从动态系统的角度来看是可解释的，并且在某种意义上它们能够将状态空间中的神经轨迹与数据中的时空活动模式相关联。因此，RNN轨迹映射到细胞组件上，如TP9中所研究的那样，从而将这两种理论方法紧密联系在一起。最后，我们将把所有的发现整合到一个共同的计算框架和前额叶灵活性背后的神经动力学原理的理论中。

The present consortium offers a unique and unprecedented opportunity to unravel neuro-dynamical and computational principles and mechanisms of higher cognitive functions across species, developmental stages, and cognitive domains. This project will exploit this opportunity and will attempt to uncover general computational principles through analyses of experimental data acquired by the other project partners from a unifying dynamical systems perspective, and within a common statistical machine learning framework. We will use deep generative recurrent neural networks (RNN) to approximate the dynamical systems underlying the experimental data. While RNNs in general have recently become a popular tool in neuroscience to study dynamical mechanisms, our methods go one step further and infer RNN parameters directly from the observed physiological and behavioral time series in a statistical, maximum-likelihood or Bayesian approach, yielding quantitative predictions. In collaboration with the experimental partners, we will deploy these methods established in our group to study the neural attractor dynamics of decision making, neuro-dynamical mechanisms underlying interval timing, species-specific neuro-computational mechanisms of working memory, and how these change with developmental stage. We will also address the long-standing question of whether similar or distinct neuro-computational principles underlie PFC-dependent performance on different cognitive tasks. In each of these instances, we propose a set of specific hypotheses about the underlying neuro-dynamical mechanisms that can be tested using our approach. Moreover, we will also generate specific (quantitative) novel predictions that will be fed back to the experimental partners. Furthermore, our RNNs are interpretable both from a dynamical systems perspective and in the sense that they enable to relate neural trajectories in state space to spatio-temporal activity patterns in the data. Hence, RNN trajectories map onto cell assemblies as studied in TP9, thus tightly linking these two theoretical approaches. Finally, we will integrate all findings into a common computational framework and theory of the neuro-dynamical principles underlying prefrontal flexibility.