"CRCNS: A Bayesian Framework for Sensorimotor Learning and Control"

描述(申请人提供)：适应是生物运动控制的核心，因为我们的肌肉，我们的运动装置，以及我们与之相互作用的环境都具有时变性。其中一些属性突然而暂时地变化，而另一些则变化缓慢，可能会持续很长一段时间。我们认为，大脑学习运动控制的方式在很大程度上反映了这些时间尺度的变化。实际上，我们认为大脑将适应视为一个统计问题，其中包括关于变化时间尺度的先验信息。我们的新方法解决了两个主要问题，这两个问题显著地限制了我们目前对适应的神经基础的理解：(1)目前的运动适应模型几乎完全集中在错误上。在每次试验中，误差都与当前环境相关联，并且系统通过改变参数值或权重来适应其内部模型。然而，有充分的证据表明，行为不仅会随着错误的大小而变化，而且会随着错误发生的时间而变化。我们的方法预测了适应实验中对训练试验的时间和上下文历史的强烈依赖。(2)传统的运动适应理论假设中枢神经系统选择一个期望的轨迹，通过试错估计内部模型参数，然后产生沿该期望轨迹移动肢体的运动指令。然而，运动的成本和收益通常被用其最终结果来描述，而不是具体的轨迹。事实上，对适应过程中的行为进行仔细检查，就会发现与预期轨迹假设的预测明显背离。在这里，我们在一个单一的框架内处理关于世界和我们的汽车工厂的不确定性问题以及成本函数。我们提出了一个新的理论来指导感觉运动学习的实验。我们在这里提出的是从根本上改变目前对运动误差和所需轨迹的关注--这些想法一直是感觉运动研究的支柱。首先，我们将内部模型的学习与人体可能如何受到现实世界扰动影响的因果结构联系起来。接下来，我们将内部模型的变化与感觉运动控制的变化联系起来。因此，我们认为，我们和其他人测量的自适应行为实际上是两个并行计算过程的结果：内部模型的统计公式(一个类似于系统识别的过程)，以及在最优控制框架中使用这些内部模型来产生运动指令。这一新理论框架的发展最终将阐明大脑中参与控制我们运动的结构所执行的计算。

DESCRIPTION (provided by applicant): Adaptation is central to biological motor control because our muscles, our motor plant, and the environment that we interact with have time varying properties. Some of these properties change suddenly and temporarily while others change slowly and may last a long time. We suggest that the way that the brain learns motor control is to a great extent a reflection of these time scales of change. In effect, we propose that the brain treats adaptation as a statistical problem that includes prior information about timescales of change. Our new approach solves two major problems which have significantly curtailed our current understanding of the neural basis of adaptation: (1) Current models of movement adaptation focus almost exclusively on error. On every trial, the error is associated with the present context and the system adapts its internal model by changing parameter values or weights. However, there is ample evidence that behavior can change not only as a function of the magnitude of errors, but also as a function of when these errors are made. Our approach predicts the strong dependence on the temporal and contextual history of the training trials in adaptation experiments. (2) Traditional motor adaptation theories assume that the CNS chooses a desired trajectory, estimates internal model parameters through trial and error, and then produces motor commands that move the limb along this desired trajectory. However, movements have costs and gains that are often described in terms of their end result, not a specific trajectory. Indeed, close inspection of behavior during adaptation reveals clear deviations from the predictions of a desired trajectory assumption. Here we treat the problems of uncertainty about the world and our motor plant as well as cost functions in a single framework. We suggest a new theory to guide experiments in sensorimotor learning. What we are proposing here is a fundamental shift away from the current focus on motor error and desired trajectories - ideas that have been the mainstays of sensorimotor research. First, we link learning of internal models to a causal structure of how the body might be affected by real world perturbations. Next, we link changes in internal models to changes in sensorimotor control. As a result, we suggest that the adaptive behavior that we and others have measured is really a result of two concurrent computational processes: statistical formulation of internal models (a process akin to system identification), and the use of those internal models in the framework of optimal control to produce motor commands. Development of this new theoretical framework will ultimately shed light on the computations that are performed by structures in the brain that participate in control of our movements.