Unexpected complexity of everyday manual behaviors

Unexpected complexity of everyday manual behaviors
Nature Communication, 2020. Chiago of University.

传统分析中，一般把大数据降维，然后用降维的数据来研究。这种做法的背景是认为降维保留了数据的大部分信息，去除了noise。但是这篇文章通过分析所谓的noise，发现这些noise仍然保留了数据的信息，证据是利用这些noise可以对动作进行正确率很高的分类。通过分析，发现noise数据包含了精细操作的信息。 而传统的降维保留数据，包含的是动作的主体运动。

这篇文章里用了两种非线性降维方法：
1）Isomap:A global geometric framework for nonlinear dimensionality reduction.
2）nonlinear DR autoencoder-based approach:Nonlinear principal component analysis: neural network models and applications.
matlab里第一种方法的package：isomap，应用时指定nearest neighbors数目。
matlab里第二种方法的package：nlpca

这篇文章里有个骚操作，需要学习：如何排除noise是condition-specific information的。首先，作者阐述说cross condition都表现这种相同的趋势，如果是condition-specific information则不可能这么高度一致。然后，作者用仿真的方法来证明：

1. 提取one trial from each object, and replicated it four more times to obtain a kinematics set that contained no within-condition noise. Then reconstructed the (denoised) kinematics with only the first 10 PCs. Next, drew from a multivariate Gaussian distribution with zero mean and a condition-specific covariance matrix, Last, randomly shuffled joint angle order and recalculated the covariance
   matrix of the denoised data, repeating this procedure for each object.。这个操作在这里就是添加具有condition-specific的noise，We then computed the PCs of these simulated kinematics and performed the same classification analysis described above with sequentially removed PCs. We repeated the procedures above five times, each time using a different seed to generate the conditional noise distributions (by reshuffling the joint angles and resampling from the resulting distributions). 进行多次重复实验。