long8v / PTIR

Paper Today I Read
19 stars 0 forks source link

[15] Quantifying Memorization Across Neural Language Models #15

Open long8v opened 2 years ago

long8v commented 2 years ago

image paper problem : 모델이 커짐에 따라 training data를 외우는 일이 생긴다. 이러한 현상이 1) 모델 크기 2) 데이터 반복 횟수 3) 주어지는 context의 길이에 따라 얼마나 증가하는지를 정량적으로 평가해본다.

conclusion : image

  1. Model scale: Within a model family, larger models memorize 2-5× more data than smaller models.
  2. Data duplication: Examples repeated more often are more likely to be extractable.
  3. Context: It is orders of magnitude easier to extract sequences when given a longer surrounding context. -> 좋은 쪽으로 해석하면 그만큼 adversarial attack을 하기 어렵다는 뜻임. Practitioners building language generation APIs could (until stronger attacks are developed) significantly reduce extraction risk by restricting the maximum prompt length available to users.

details :

next papers :

thinkings :