Traditional neural network input is d-dimensional, is an m-dimensional output, i.e. d * m w is the dimension. Now with maxout, w is the dimension d * m * k. It is now an m-dimensional output, but before the output, for each of the m nodes has a dimension k, k to obtain a maximum of the selected node m.
https://blog.csdn.net/hjimce/article/details/50414467 This article has a good explanation.