In this paper, we propose an idea to improve various types of loss functions. It is different from the current idea of balancing the errors by increasing the number of input samples in each batch. We directly mask the top values of the error ranking to zero. During the error back-propagation, this means that the samples corresponding to that loss will not affect the parameter update of the network. In other words, even if a small number of samples are artificially mislabeled, it will not theoretically have much impact on the performance of the network. Instead, deliberately discarding anomalous losses will help smooth the training of the network. We conduct experiments on several regression and classification tasks, and the results show that the proposed method in this paper can effectively improve the expected performance of the network.

In this paper, we propose a new clustering module that can be trained jointly with existing neural network layers.  Specifically, we have designed a generic clustering module with a competitive update mechanism. The module consists of a Gaussian unit and a maximum pooling layer. The Gaussian unit forward propagation conforms to the joint Gaussian distribution and contains two sets of trainable parameters. It requires no tedious setup and has a plug-and-play feature. To improve the  representation capability of the network, we used an auto-encoder to extract the hidden semantics of the input features and combined the clustering module with the auto-encoder to construct an end-to-end unsupervised clustering neural network. We refer to this as HGL_CAE(High-dimensional Gaussian distribution layers combined with convolutional autoencoders).The network is highly adaptable to different input feature dimensions and can cope with situations where the number of clusters  cannot be determined in advance. We conducted experiments on the MNIST dataset and the Fashion_MNIST dataset with a clustering accuracy of 93.38% and 72.83% respectively. It is highly competitive with existing methods.