Learning sensorimotor maps through curiosity

想象你是一个第一次学习说话或演奏乐器的孩子。你的大脑需要知道哪个动作会产生哪个声音。运动神经科学的一个难题是运动系统如何解决这一挑战：我们如何将我们所做的（运动）与我们所看到或听到的（感知）联系起来，也就是说，我们如何学习感觉运动地图。我的研究项目的长期目标是提供一种新的方式来思考这个由信息寻求驱动的过程：好奇心。学习感觉运动图可以用我之前介绍过的一个实验范式来研究，在这个实验范式中，参与者的手臂运动被映射到声音上。由于手臂运动通常不会产生声音，这种模式允许我们在学习的第一阶段监测参与者，在那里他们形成了一个新的感觉运动地图。现在我们可以问是什么驱动了我们观察到的学习。迄今为止使用这一范式的工作表明，当受试者被允许自由运动时，他们可以获得相应的反馈声音，学习就会发生。随后的测试表明受试者获得了地图。这表明，学习是由一种不依赖于目标的原则驱动的。这个DG循环的重点是一个假设，即地图学习是由对信息的搜索驱动的，也就是好奇心。运动神经科学先前的工作研究了随机探索，而在机器学习算法领域，已经开发出基于对自己预测准确性的内部估计，以原则性的方式探索其环境的算法。这种好奇心的概念还没有被应用到人类运动学习的研究中，而这正是我们将要做的。目的：1)人为操纵映射的可预测性，以测试信息寻找是否构成学习的基础。2)我们将采用机器学习中的算法来开发一种新的、基于好奇心的人类如何学习感觉运动地图的模型。我们将测试这个模型是否能比其他模型更好地预测人类行为，这些模型不包含好奇心，而是随机探索，如果是这样的话，将表明好奇心推动人类地图学习。3)我们将测试人类运动网络中的大脑区域参与基于好奇心的学习的假设。我们将使用磁共振波谱（MRS），一种可以测量大脑中兴奋性和抑制性神经化学物质浓度的技术，来收集在扫描仪中执行学习任务的人类受试者的数据。我们将跟踪每个动作的预期信息增益，从而形成预期活动的精确模式。这种模式可以与来自不同大脑区域的观察数据进行比较，从而确定基于好奇心的学习的神经基础。综上所述，这些研究将为运动学习探索阶段的基础提供急需的见解。

Imagine you are a child first learning to speak or to play a musical instrument. Your brain will need to learn which movement leads to which sound. One of the puzzles of movement neuroscience is how the motor system solves this challenge: how we link what we do (movement) with what we see or hear (perception), that is, how we learn a sensorimotor map. The long term aim of my research program is to provide a new way of thinking about this process as driven by information seeking: curiosity. Learning sensorimotor maps can be studied using an experimental paradigm that I introduced previously in which participants make arm movements that are mapped to sounds. Since arm movements normally do not result in sounds, this paradigm allows us to monitor participants in the first stages of learning where they form a novel sensorimotor map. Now we can ask what drives the learning we observe. Work to date using this paradigm has revealed that learning can occur when subjects are allowed to make free movements, for which they are given the corresponding feedback sounds. Subsequent tests show that subjects acquired the map. This suggests that learning is driven by a principle that does not rely on targets. The focus for this DG cycle is the hypothesis that map learning is driven by the search for information, that is, curiosity. Prior work in motor neuroscience has investigated random exploration, whereas in the field of machine learning algorithms have been developed that explore their environment in a principled way, based on an internal estimation of the accuracy of their own predictions. This notion of curiosity has not been applied to the study of human motor learning which is what we will do here. Aims: 1) We artificially manipulate the predictability of the mapping in order to test whether information seeking forms the basis of learning. 2) We will adapt algorithms from machine learning to develop a new, curiosity-based model of how humans learn sensorimotor maps. We will test whether this model can predict human behavior better than other models that do not incorporate curiosity but rather random exploration, which, if the case, will show that curiosity drives human map learning. 3) We will test the hypothesis that brain areas in the human motor network are involved in curiosity-based learning. We will use magnetic resonance spectroscopy (MRS), a technique that allows measuring the concentration of excitatory and inhibitory neurochemicals in the brain, to collect data from human subjects who perform our learning task in the scanner. We will track the expected information gain of each movement, thus forming a precise pattern of expected activity. This pattern can be compared against the observed data from various brain areas, thus identifying the neural basis of curiosity-based learning. Taken together, these studies will provide much needed insight into the basis of exploratory phases of motor learning.