Version mismatch in embedding-based retrieval

July 31, 2021 by admin

Version mismatch in embedding-based retrieval is challenging, esp. on the infra side.

https://recsysml.substack.com/p/a-common-mistake-when-using-embeddings

Shujian Follow

Software engineer @Google Travel. PhD in renewable energy from @UMassAmherst. @kaggle triple master. IR/NLP/ASR.

Retweet on Twitter Shujian Retweeted

François Chollet @fchollet ·

18 Mar

"It's autocomplete" is not a helpful analogy to understand LLMs. A LLM is more like a database that lets query information in natural language. You can query both knowledge, and "patterns" (associative programs seen in the training data, that can be applied to new inputs).

Reply on Twitter 1637121320340299776 Retweet on Twitter 1637121320340299776 157 Like on Twitter 1637121320340299776 1221 Twitter 1637121320340299776

Retweet on Twitter Shujian Retweeted

Jason Wei @_jasonwei ·

13 Mar

Hot take supported by evidence: for a given NLP task, it is unwise to extrapolate performance to larger models because emergence can occur.

I manually examined all 202 tasks in BIG-Bench, and the most common category was for the scaling behavior to *unpredictably* increase.

Reply on Twitter 1635338409370865665 Retweet on Twitter 1635338409370865665 56 Like on Twitter 1635338409370865665 359 Twitter 1635338409370865665

Retweet on Twitter Shujian Retweeted

Cosmin Negruseri @cosminnegruseri ·

28 Feb

this slide is great, and focuses on one ranking model + postprocessing, but if your team owns an end to end system with indexing, candidate retrieval, ranking, blending oncall is even more complicated

Reply on Twitter 1630451840503668740 Retweet on Twitter 1630451840503668740 3 Like on Twitter 1630451840503668740 17 Twitter 1630451840503668740

Retweet on Twitter Shujian Retweeted

Jo Kristian Bergum @jobergum ·

27 Feb

Hm, a ready-to-ship e-commerce search solution with tunable hybrid ranking, auto-complete query suggestions, and query contextualized navigation. Better than any commercial vendor, but with open-source technology, seeing how the sausage is made.

Reply on Twitter 1630275264314744833 Retweet on Twitter 1630275264314744833 7 Like on Twitter 1630275264314744833 81 Twitter 1630275264314744833

Retweet on Twitter Shujian Retweeted

Zhiqing Sun @edwardsun0909 ·

22 Feb

How can LLMs such as GPT-3 and ChatGPT achieve greater factual accuracy without relying on an external retrieval search engine?

Our #ICLR2023 paper shows that recitation can help - like humans!

Recitation-Augmented Language Models
https://arxiv.org/abs/2210.01296

1/N

Reply on Twitter 1628494281588740096 Retweet on Twitter 1628494281588740096 92 Like on Twitter 1628494281588740096 376 Twitter 1628494281588740096