Recent work on robustness againt adversarial examples identified a severe problem in adversarial training: (robust) overfitting. That is, during training the training robustness continuously increases, while test robustness starts decreasing eventually. In this pre-print, we relate robust overfitting and good robust generalization to flatness around the found minimum in the robust loss landscape with respect to perturbations in the weights.
In this MLSys’21 paper, we consider the robustness of deep neural networks (DNN) against bit errors in their quantized weights. This is relevant in the context of DNN accelerators, i.e., specialized hardware for DNN inference: In order to reduce energy consumption, the accelerator’s memory may be operated at very low voltages. However, this induces exponentially increasing rates of bit errors that directly affect the DNN weights, reducing accuracy significantly. We propose a robust fixed-point quantization scheme, weight clipping as regularization during training and random bit error training to improve bit error robustness. This article shares my talk recorded for MLSys’21.