Correction of erroneously-trained adaptive neural networks

Abstract

Applications of adaptive neural network models, i.e. models that are updated while they are used, are becoming more widespread. Adjustments to an adaptive model are based on a feedback which is collected while the model is in operation. While changing model parameters in production can improve the performance, itcomes with inherent risks. In particular, any erroneous adjustment to an online model may reduce the performanceand be detrimental to control applications and costly to the business. At the same time, mistakes in feedback are sometimes unavoidable. Therefore, an adaptively trained model with old data is sub-optimal unless the new revisions are taken into account. In this paper, we investigate the effect of a faulty feedback on the performance of an e-commerce customer identification neural network model. We first investigate the impact of feedback error on an adaptive model's performance. We then examine a technique to undo the incorrect adjustments to the model by re-training the adaptive model by a corrected feedback. Our results showthe majority of lossinmodel performance due to the feedback error is recoveredby re-training the adaptive model with the new corrected data.