RETRO

姚伟峰(Matrix Yao)

• RETRO
  • Info Card
  • Basic Idea
  • How Does it Work
    • Step-1: Retrieve Nearest Neighbors and Encode them
    • Step-2: Decode Causally
  • Results
    • Language Model
    • Downstream Task: QA
  • Application on ODQA domain
    • Pipeline Comparison
    • To put RETRO into the map of ODQA paradigms
  • References

Info Card

• full name
  Retrieval Enhanced TRansfOrmer

• paper
  Improving language models by retrieving from trillions of tokens

• year
  2022

• from
  

• GitHub
  un-official

Basic Idea

RETRO is a neural language model.
Comparing w/ existing language models like GPT, it separates memorization and generalization, so memorize the world knowledge w/ Retrieval, while learn the language structure w/ Model.

General auto-regressive language model

RETRO’s chunked retrieval enhanced model

 

• Any benefits?
  Democratization fast/cheap and good

  • Fewer parameters: 25x fewer parameters lead to much lower requirement computation requirement for training and serving;

  • SOTA accuracy: show better perplexity on LM and SOTA accuracy on downstream tasks e.g., question answering;

Below diagram from [1] is not the whole picture of RETRO. It’s just the retrieval part.

How Does it Work

Step-1: Retrieve Nearest Neighbors and Encode them

• Points

  • retrieve the top-k nearest neighbors in chunk granularity, neither passage granularity as sentence BERT nor token granularity like ColBERT

  • each of top-k token sequence = concat(neighbor chunk, continuation chunk)

  • each token sequence is encoded w/ bi-directional transformer encoder, optionally w/ self-attended query as k/v

Step-2: Decode Causally

• CCA(Chunked Cross Attention)
  

• Points

  • both attention and CCA are causal, to make it auto-regressive

Results

Language Model

Pretty good bits-per-byte even 23+x smaller size.

Downstream Task: QA

Not really so good, considering the 7.5B model size. And inferior accuracy than FiD, they blame the encoder weight not enough in current model.

Application on ODQA domain

Pipeline Comparison

• dense retriever + neural ranker
  E.g.,

  • Single Retrieval Encoder: SentenceEmbedding Retriever + ColBERT Ranker

  • Dual Retrieval Encoder: DPR Retriever + ColBERT Ranker

  

• RETRO
  

We can see that RETRO can easily fit as a dense retriever + neural ranker ODQA pipeline. It can be viewed as single-encoder dense retriever + neural ranker , and the ranker is compute-heavier than ColBERT, both because of model size and the ranker doc encoder cannot be pre-computed.

To put RETRO into the map of ODQA paradigms

References

1. RETRO Is Blazingly Fast

2. The Illustrated Retrieval Transformer

标签：ranker,language,RETRO,model,ODQA,Retrieval
来源： https://www.cnblogs.com/Matrix_Yao/p/16480698.html