Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

ALBERT: A LITE BERT FOR SELF-SUPERVISED LEARNING OF LANGUAGE REPRESENTATIONS(ICLR 2020) #9

Open
kirinsannnnnnnnnn opened this issue Oct 8, 2019 · 0 comments
Labels
2020 ICLR NLP Natural Language Processing Yuta Tsuchizawa

Comments

@kirinsannnnnnnnnn
Copy link
Collaborator

matome information

  • matome author: Yuta Tsuchizawa
  • read date: 20191008

ALBERT: A LITE BERT FOR SELF-SUPERVISED LEARNING OF LANGUAGE REPRESENTATIONS

paper information

1. What did authors try to accomplish?

  • BERTを効率よく学習できるように軽量化し、さらに精度改善を試みた

2. What were the key elements of the approach?

  • 軽量化・高速化
    1. factorizing embedding parameterization
    2. cross-layer parameter sharing
    • 結果、trainingがパラメータが18倍少なくなり1.7倍高速化した
  • 精度改善
    1. (NSP)Next Sentence Prediction → (SOP)Sentence Order Prediction
    • w/self-supervised loss
  • 評価
    • GLUE, SQuAD, RACEでsota

3. What can you use yourself?

  • 単純に一回試してみたい

4. What other references do you want to follow?

その他メモ

github

official blog

  • なし

author's resource

  • なし

third person's explanation

感想

  • Cross-Layer Parameter SharingとFactorized Embeddingで永続化する必要があるパラメータ数は10~30倍減ったのはすごい
    • ALBERT xlargeでBERT baseよりパラメータが少なく精度が高いので試す価値は十分ありそう
  • 一方で計算時間は2倍程度早くなっている
    • 本番運用する上ではこちらも重要なので今後の高速化に期待したい

分野全体での立ち位置

  • BERTを改善しました系
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
2020 ICLR NLP Natural Language Processing Yuta Tsuchizawa
Projects
None yet
Development

No branches or pull requests

1 participant