You Only Sample (Almost) Once: Linear Cost Self-Attention Via Bernoulli Sampling

Benchmark Model Rank Results
paraphrase-identification-on-mrpc-glue-devYOSO-E#9F1: 88.1
natural-language-inference-on-qnli-glue-devYOSO-E#14Accuracy: 90.1
natural-language-inference-on-multinli-devYOSO-E (MNLI-m)#18Matched: 82.2
sentiment-analysis-on-sst-2-binary-classificationYOSO-E#92Dev Accuracy: 92.3