Competitive Cross-Entropy Loss: A Study on Training Single-Layer Neural Networks for Solving Nonlinearly Separable Classification Problems

摘要

After Minsky and Papert (Perceptrons, MIT Press, Cambridge, 1969) showed the inability of perceptrons in solving nonlinearly separable problems, for several decades people misinterpreted it as an inherent weakness that is common to all single-layer neural networks. The introduction of the backpropagation algorithm reinforced this misinterpretation as its success in solving nonlinearly separable problems passed through the training of multilayer neural networks. Recently, Conaway and Kurtz (Neural Comput 29(3):861–866, 2017) proposed a single-layer network in which the number of output units for each class is the same as input units and showed that it could solve some nonlinearly separable problems. They used the MSE (Mean Square Error) between the input units and the output units of the actual class as the objective function for training the network. They showed that their method could solve the XOR and M&S’81 problems, but it could not do any better than random guessing on the 3-bit parity problem. In this paper, we use a soft competitive approach to generalize the CE (Cross-Entropy) loss, which is a widely accepted criterion for multiclass classification, to networks that have several output units for each class, calling the resulting measure the CCE (Competitive cross-entropy) loss. In contrast to Conaway and Kurtz (2017), in our method, the number of output units for each class can be chosen arbitrarily. We show that the proposed method can successfully solve the 3-bit parity problem, in addition to the XOR and M&S’81 problems. Furthermore, we perform experiments on several datasets for multiclass classification, comparing a single-layer network trained with the proposed CCE loss against LVQ, linear SVM, a single-layer network trained with the CE loss, and the method of Conaway and Kurtz (2017). The results show that the CCE loss performs remarkably better than existing algorithms for training single-layer neural networks.