Remember to correct the bias when using deep learning for regression!
This solves a practical problem for practitioners using deep learning for regression by preventing error accumulation in aggregated results, though it is incremental as it builds on existing training methods.
The paper addresses the systematic error in deep learning regression models where training residuals don't sum to zero, leading to accumulated errors in aggregated performance, and proposes a bias correction postprocessing step that effectively mitigates this issue, as shown in experiments.
When training deep learning models for least-squares regression, we cannot expect that the training error residuals of the final model, selected after a fixed training time or based on performance on a hold-out data set, sum to zero. This can introduce a systematic error that accumulates if we are interested in the total aggregated performance over many data points. We suggest to adjust the bias of the machine learning model after training as a default postprocessing step, which efficiently solves the problem. The severeness of the error accumulation and the effectiveness of the bias correction is demonstrated in exemplary experiments.