Computational mechanisms of the head direction circuit development shaped by visuomotor experience

到目前为止，我提出了一个头部方向表征的无监督学习框架，并将其作为机器人系统在定制的虚拟物理环境中实现；首先，我建立了一个机器人平台，用于数据记录和学习算法的训练和测试；其次，我开发了一个无监督深度神经网络，它可以在训练过程中自动形成头部方向表征；第三，我测试了无监督头部方向神经网络的性能。预测的头部方向与地面真实之间的误差在-21度至27度之间，均方根误差约为11.1度。4度排列的磁头方向单元的调谐曲线显示一个周期性的凸起来表示磁头方向。

Until now, I proposed a framework for unsupervised learning of head direction representations, which is implemented as a robotic system in customized virtual physical environments.First, I set up a robotic platform for data recording and learning algorithm training and testing.Second, I developed an unsupervised deep neural network, which can automatically form head direction representations during the training process.Third, I tested the performance of unsupervised head direction neural networks. The error between the predicted head direction and ground truth is between -21 degrees and 27 degrees with a root mean square error of about 11.1 degrees. The tuning curves of head direction cells binned by 4 degrees show a periodic bump to represent the head direction.