这个在autoencoder与PCA比较时，有人做过类似的工作吧。
他是假设weight matrix的augmented eigenvalues是单位矩阵，所以panelty多了一项w*w'-I。
【 在 Charles9 的大作中提到: 】
: Any researching on keeping Weights linear independent during training? I am thinking that no matter how you transform features on high dimensional space, if these features are independent, the vector representation should also be linear independent and these vectors span the problem space.
: 发自「今日水木 on iPhone 8」
--
FROM 117.136.38.*