Skip to main content
. 2021 Jul 5;15:675741. doi: 10.3389/fncom.2021.675741

Figure 5.

Figure 5

(A) Without any noise-aware techniques, inference on the Microsoft Research Paraphrase Corpus (MRPC) task decays very quickly over time. (B) Drift compensation improves the decay over time significantly, but the inference results are still lower than the BERT-base ideal model with no NVM noise. (C) Hardware-aware (HWA) training with noise added during training helps close the gap, reaching software-equivalent accuracy for this task even at 1 month of drift.