Neural classifiers using loss functions

Summary form only given. While a vast majority of backpropagation-based classifiers use mean squared error (MSE) as an error measure, it can be shown that MSE is inadequate for classification for three reasons: (1) its minimum is different from the minimum of misclassification loss, (2) it is unable to account for class-specific misclassification losses, and (3) it slows learning by imposing unnecessary constraints. By contrast, using the misclassification loss as error measure overcomes all these problems. Its use with backpropagation is as easy as that of MSE. Computational experiments with a differentiable approximation of the misclassification loss have confirmed its superiority over the MSE in terms of both convergence speed and misclassification rates