Statistical machine learning tools for understanding neural ensemble representations and dynamics

大脑是一个由大量相互连接的专门电路组成的网络。了解这些电路如何支持 感觉、感知、认知和行动需要测量区域内和区域间的活动模式，但是 测量本身并不能洞察潜在神经元的结构或功能。 系统。洞察力需要应用量化方法，将神经元活动模式与 可通过实验测量的变量，包括现在和过去的感觉输入、当前位置和 当前或未来的电机输出。其结果是建立了一个将测量变量与峰值活动联系起来的“编码”模型。 通过简单地应用贝叶斯规则，这种编码模型可以被用来创建一种“解码”模型。在……里面 解码，目标是采用尖峰活动的模式，以及之前开发的编码模型，以及 评估与尖峰相对应的感觉、认知或运动表征。编码和解码 算法是现代系统神经科学的基本组成部分，在帮助我们 了解神经元表征的本质和动力学。这些方法提供了一种强大的方法来 深入了解神经元群体，但当前算法的几个局限性削弱了它们的有效性。第一, 虽然现代深度神经网络在解码方面可能是强大的，但它们在这方面有多个缺点 科学发现的最高成就。其次，高级译码算法往往过于复杂且计算量太大 对于大多数研究人员来说，在大规模神经数据集的分析中实现是密集的。此外，健壮， 易于使用的软件允许不太老练的用户利用这些算法，但不能 是存在的。第三，解码的结果通常对记录的神经元总数非常敏感。第四, 同时解码单个变量(例如，动物位置、目标值等)很容易处理，解码多个变量 同时，这超出了当前方法的能力范围。第五，神经反应的性质和质量 神经记录的频率在实验过程中经常会发生变化。现有的解码算法要么是 静态或需要重复重新估计编码模型以保持估计精度。最后，解码 传统上专注于观察到的信号，如动物的位置，但最近的研究重点是 未观察到的认知过程，如心理探索。需要新的方法来确定何时 认知过程的解码是可靠的。解决这些问题需要新的方法和新的 使这些方法易于使用且对社区高效的并行软件。我们有 开发的无簇解码算法非常有效地利用了可用的数据，在这里我们将 进一步开发这些算法和实现这些算法的软件，以应对所述的所有挑战 上面。其结果将是一套强大的工具，有可能推动新的发现。

The brain is a massively interconnected network of specialized circuits. Understanding how these circuits support sensation, perception, cognition, and action requires measuring activity patterns within and across regions, but the measurements themselves do not produce insight into the structure or function of the underlying neuronal system. Insight requires the applications of quantitative methods that relate neuronal activity patterns to experimentally measurable variables, including things like present and past sensory inputs, current location, and current or future motor outputs. The result is an “encoding” model relating measured variables to spiking activity. Through a simple application of Bayes rule, this encoding model can be used to create a “decoding” model. In decoding, the goal is to take a pattern of spiking activity, along with a previously developed encoding model, and assess the sensory, cognitive or motor representation corresponding to the spiking. Encoding and decoding algorithms are a fundamental part of modern systems neuroscience and play a critical role in helping us understand the nature and dynamics of neuronal representations. These approaches provide a powerful way to gain insight about neuronal populations, but several limitations of current algorithms blunt their efficacy. First, while modern deep neural networks can be powerful for decoding, they have multiple shortcomings in the context of scientific discovery. Second, advanced decoding algorithms tend to be too complex and computationally intensive for most researchers to implement in the analyses of large-scale neural datasets. Moreover, robust, easy to use software that would allow less sophisticated users to take advantage of these algorithms does not exist. Third, the results of decoding are typically very sensitive to the total number of neurons recorded. Fourth, while decoding a single variable (e.g. animal position, target value, etc.) is tractable, decoding multiple variables simultaneously is beyond the capacities of current approaches. Fifth, neural response properties and the quality of neural recording often changes through the course of an experiment. Existing decoding algorithms are either static or require repeated re-estimation of the encoding model to maintain estimation accuracy. Finally, decoding has traditionally focused on observable signals, such as the animal’s position, but recent work has focused on unobserved cognitive processes, such as mental exploration. New methods are needed to determine when decoding of cognitive processes is reliable. Solving these problems requires new approaches and new parallelized software that make these approaches easy to use and efficient for the community. We have developed clusterless decoding algorithms that make very efficient use of the available data, and here we will further develop those algorithms and the software that implements them to meet all of the challenges described above. The result will be a powerful set of tools that have the potential to drive new discoveries.