Prospective coding by cortical pyramidal neurons

这项牵头机构的提案是W.Senn关于树枝计算理论的个人SNF-GRANT的延续。在运行项目期间，一个关键的洞察力表明，在神经元水平上的学习可以预测树突输入的体细胞尖峰。到目前为止，一般的实验和理论研究都试图证明突触信号的树突加工过程中存在复杂的非线性。但是，仅仅将神经元描述为复杂的输入输出元素，对于树突实际计算的内容几乎没有什么帮助。相反，将神经元视为内在预测元素将单个神经元的处理与可能的更广泛的计算任务联系起来。当前的提议通过展望编码的概念扩展了这一单个神经元的假设。这一概念意味着神经元的活动预测了当前以及未来的突触输入。我们假设，锥体神经元的基础树和顶树都基于与基础树的网络内突触和与顶树的外部突触，对体细胞尖峰进行独立的预测。独立预测之间的匹配代表着一个高度可信的信号，它产生树突状钙峰和随后的躯体动作电位爆发，然后可以反馈给突触前神经元，后者可以将它们用作自己上游突触的教学信号。本文的理论和实验验证分为四个子项目：SP1：前瞻性编码(牵头：SENN实验室)。形式化展望编码的概念，并证明基础树和树状树对未来输入的独立预测等价于贝叶斯线索组合问题。SP2：时间反向传播(Lead：Senn Lab)。表明用于预测未来事件的匹配信号可以用于训练对这些预测有贡献的隐藏神经元。将该理论应用于非马尔科夫序列学习问题和简化的接球问题。SP3：新颖性编码(领导：拉库姆实验室)。在体内测试树枝状钙尖峰是否代表预测信号之间的匹配，或者是基础树和顶端树产生的新奇信号之间的匹配。通过结合听觉和躯体感觉线索测量神经元对躯体感觉古怪范式的反应，验证线索组合假说的预测。SP4：纠错可塑性(Lead：Neian Lab)。体外验证假设，突触可塑性在兴奋性和抑制性可塑性中都是纠错的，因此是否是非Hebbian的。测试涉及钙离子尖峰的可塑性是否如预期编码所预测的那样具有较长的诱导时间窗口。前两个子项目将产生正式框架，随后的两个实验子项目将嵌入其中。它们将被联合设计，结果将以数学模型的形式描述。

This Lead Agency proposal is a continuation of the personal SNF-grant of W. Senn on the theory of dendritic computation. In the running project period a key insight has suggested that learning on the level of a neuron predicts somatic spiking by dendritic inputs. So far, general experimental and theoretical research has tried to prove complex nonlinearities in the dendritic processing of synaptic signals. But the mere description of a neuron as a complex input-output element gives only little insight into what dendrites are actually computing. In contrast, regarding neurons as intrinsic prediction elements links single neuron processing to a possible broader computational task.The current proposal extends this single neuron hypothesis by the notion of prospectivecoding. This notion implies that the activity of a neuron predicts current, as well as future synaptic inputs. We hypothesize that both the basal and apical dendritic tree of a pyramidal neuron make independent predictions of the somatic spiking, based on within-network synapses to the basal tree and extrinsic synapses to the apical tree. The match between the independent predictions represents a high confidence signal that generates dendritic calcium spikes with a subsequent burst of somatic action potentials, which can then be fed back to the presynaptic neurons that can use them as a teaching signal for their own up-stream synapses. The theory and its experimental verification is divided into 4 subprojects:SP1: Prospective coding (Lead: Senn lab). Formalize the concept of prospectivecoding and show that the independent prediction of future input by the basal and dendritic trees is equivalent to a Bayesian cue combination problem.SP2: Backpropagation in time (Lead: Senn lab). Show that the matching signal for the prediction of future events can be used to train hidden neurons that contribute to these predictions. Apply the theory to the non-Markovian sequence learning problem and to a simplified ball catching problem.SP3: Novelty coding (Lead: Larkum lab). Test in vivo whether a dendritic calcium spike is representing the match between prediction signals or between novelty signals generated by the basal and apical trees. Verify the prediction of the cue combination hypothesis by measuring the neuronal responses to a somatosensory oddball paradigm with combined auditory and somatosensory cues.SP4: Error-correcting plasticity (Lead: Nevian lab). Verify the hypothesis invitro whether synaptic plasticity both in excitatory and inhibitory plasticity is error-correcting and hence non-Hebbian. Test whether plasticity involving calcium spikes has a longer induction time window as predicted by prospective coding.The first two subprojects will yield the formal framework in which the subsequent two experimental subprojects are embedded. They will be jointly designed and the results will be described in terms of a mathematical model.