Confidence estimation method for regression neural networks

Abstract

Numerous confidence estimation methods have been proposed for classification neural networks; however, this problem has not been well addressed for regression neural networks. That is, softmax layers are not available in regression networks and the interpretation of confidence becomes less clear. To alleviate these problems, a simple but effective method is proposed that computes the confidences of regression results. First, the confidence is considered as a scalar value representing relative error-levels. Then, a mini-batch based training method based on this interpretation is developed. Precisely, in each mini-batch, desired outputs for confidence values are assigned by sorting current errors. Experimental results on the loose wheel nut detection problem as well as a simulated example have shown that the proposed method can be successfully applied to regression problems.