Global quantitative robustness of regression feed-forward neural networks
This work addresses the robustness of neural networks for regression tasks, which is an incremental contribution by applying classical robust statistics to a modern context.
The authors adapted the regression breakdown point concept to feed-forward neural networks and computed breakdown points for various network configurations and contamination settings, finding that robust loss functions improve performance in terms of out-of-sample loss, breakdown rate proxy, and training steps.
Neural networks are an indispensable model class for many complex learning tasks. Despite the popularity and importance of neural networks and many different established techniques from literature for stabilization and robustification of the training, the classical concepts from robust statistics have rarely been considered so far in the context of neural networks. Therefore, we adapt the notion of the regression breakdown point to regression neural networks and compute the breakdown point for different feed-forward network configurations and contamination settings. In an extensive simulation study, we compare the performance, measured by the out-of-sample loss, by a proxy of the breakdown rate and by the training steps, of non-robust and robust regression feed-forward neural networks in a plethora of different configurations. The results indeed motivate to use robust loss functions for neural network training.