Orthogonal Regularization

正交性在ConvNet过滤器中是一种理想的品质,部分原因是与正交矩阵相乘会使原始矩阵的norm保持不变。这一特性在深度或循环网络中很有价值,因为重复的矩阵乘法会导致信号消失或爆炸。我们注意到用正交矩阵初始化权重的成功(Saxe等人,2014),并认为在整个训练中保持正交性也是可取的。为此,我们提出了一种简单的权重正则化技术,即正交正则化,通过将权重推向最近的正交流形来鼓励权重的正交。(we propose a simple weight regularization technique, Orthogonal Regularization, that encourages weights to be orthogonal by pushing them towards the nearest orthogonal manifold.)
在这里插入图片描述

Where Σ indicates a sum across all filter banks, W is a filter bank, and I is the identity matrix.

猜你喜欢

转载自blog.csdn.net/weixin_37958272/article/details/120176289