A dendritic nexus in the circuits that coordinate learning

摘要 当我们学习一个复杂的行为时，神经系统必须不断地驱动新的动作， 预测行为对结果的影响，并加强或削弱神经元之间的连接 （突触），以改善未来的行动。然而，在控制行为的多层大脑网络中， 修改突触的行为影响取决于许多下游连接。因此，学习 这需要大脑解决一个“信用分配”问题：关于哪些突触修改应该是 made分布在整个网络中，但必须以某种方式被本地流程利用， 单个突触。我们将行为事件与突触变化联系起来的能力的一个主要差距是目前缺乏 了解这些引导单个神经元突触变化的局部过程。最近的理论 学习表明，在皮层神经元的顶端树突中产生的尖峰可能在解决问题中发挥关键作用。 这个学分分配问题在这个提议中的实验将测试假设， 前运动皮层的神经元整合了多种学习指导反馈源， 适当的条件-产生树突棘波，迅速重新配置连接和功能， 神经元在这些实验中，我们将使用先进的光学技术来监测和操纵活动， 额叶皮层中神经元子集的树突，在行动计划中具有明确的作用。一个关键 我们假设的预测是，顶端树突的活动反映了局部信用相关的计算， 这种活动不同于通过在神经元附近产生动作电位传递到其他神经元的活动。 细胞体我们将在学习过程中使用皮质神经元的纵向双光子钙成像来测试这一点， 确定树突和细胞体的行为选择性如何随行为变化而变化。为了 为了确定树突棘波对学习的贡献，我们还将使用光遗传学选择性地抑制 在学习过程中顶端树突的活动。计算模型还预测， 由来自长期反馈预测的结果信息和来自当地的 预测这种反馈的抑制。为了验证这一点，我们将结合联合收割机突触谷氨酸成像和光遗传学 将反馈投射的选择性和解剖学特性映射到顶端树突， 以确定靶向顶端树突的局部抑制性神经元的选择性。这些研究一起 将提供关键的新见解的电路机制，支配皮层可塑性和信用分配。 在这样做的过程中，他们将提供一个关键的框架，将复杂的学习与个人的修改联系起来 突触水平，并将建立机器学习算法和生物神经模型之间的桥梁 网络.

Abstract When we learn a complex behavior the nervous system must continuously drive new actions, compare predictions for the actions against outcomes, and strengthen or weaken the connections between neurons (synapses) in order to improve future actions. However, within the multilayer brain networks that control behavior, the behavioral impact of modifying a synapse depends upon many downstream connections. Thus, learning requires the brain solve a ‘credit assignment’ problem: information about which synaptic modifications should be made is distributed across the network, yet must somehow be leveraged by local processes to guide change at individual synapses. A major gap in our ability to relate behavioral events to synaptic change is the current lack of knowledge of these local processes that guide synaptic changes at individual neurons. Recent theories of learning suggest that spikes generated in the apical dendrites of cortical neurons may play a key role in solving this credit assignment problem. The experiments in this proposal will test the hypothesis that the apical dendrites of neurons in the pre-motor cortex integrate multiple learning-instructive feedback sources, and – under appropriate conditions – generate dendritic spikes that rapidly reconfigure the connectivity and function of neurons. In these experiments we will use advanced optical techniques to monitor and manipulate activity in the dendrites of a subset of neurons in the frontal cortex that have a well-delineated role in action planning. A key prediction of our hypothesis is that the activity of the apical dendrites reflects local credit-related calculations and that this activity is distinct from the activity transmitted to other neurons by action potential generation near the cell body. We will test this using longitudinal two-photon calcium imaging of cortical neurons during learning to determine how the behavioral selectivity of dendrites and cell bodies change with changing behavior. In order to identify the contribution of dendritic spikes to learning, we will also use optogenetics to selectively suppress activity in the apical dendrites during learning. Computational models also predict that dendritic spikes are generated by a mismatch between outcome information arriving from long-range feedback projections and local inhibition that predicts this feedback. To test this, we will combine synaptic glutamate imaging and optogenetics to map the selectivity and anatomical identity of feedback projections to the apical dendrites, and calcium imaging to determine the selectivity of local inhibitory neurons that target the apical dendrites. Together, these studies will provide critical new insights into the circuit mechanisms governing cortical plasticity and credit assignment. In doing so, they will provide a key framework for connecting complex learning with modifications at the individual synapse level, and will build bridges between machine learning algorithms and models of biological neural networks.