Julius Cheng's Avatar

Julius Cheng

@juliuscheng.bsky.social

Finishing up PhD in NLP at University of Cambridge. Deciding whether to put my weirdo ML thoughts on here or just be normal

24 Followers  |  30 Following  |  7 Posts  |  Joined: 23.01.2025  |  1.6106

Latest posts by juliuscheng.bsky.social on Bluesky

They say it's something like 20-30% but 0% of my papers get accepted!! Something definitely wrong here

23.01.2025 11:05 β€” πŸ‘ 1    πŸ” 0    πŸ’¬ 0    πŸ“Œ 0

The paper comes with mysterious fig 1. πŸ‘οΈ

23.01.2025 01:50 β€” πŸ‘ 9    πŸ” 1    πŸ’¬ 0    πŸ“Œ 0

Our experiments are on machine translation, but this method works with any generator + reranker setup!

Eager to hear your thoughts, and happy reranking!

23.01.2025 01:32 β€” πŸ‘ 2    πŸ” 0    πŸ’¬ 0    πŸ“Œ 0
Post image

Bonus: we show how to use multi-fidelity Bayesian optimization to use a smaller and faster proxy scoring model to search even more efficiently. We get the best performance by training a distilled model from our main CometKiwi model.

23.01.2025 01:32 β€” πŸ‘ 2    πŸ” 0    πŸ’¬ 1    πŸ“Œ 0
Post image

The candidate pool is actually a search space, and you can model your uncertainty about scores you haven't scored yet with GP regression. Use BayesOpt to search the pool for promising candidates.

This nearly gets the maximum achievable score with only 70/200 scoring calls!

23.01.2025 01:32 β€” πŸ‘ 1    πŸ” 0    πŸ’¬ 1    πŸ“Œ 0

Reranking is expensive and we show that you don't need to score every candidate in the candidate pool.

Use Bayesian optimization with GPs!

23.01.2025 01:32 β€” πŸ‘ 1    πŸ” 0    πŸ’¬ 1    πŸ“Œ 0

Language models for MT are good at generating large candidate pools that contain good translations; they're less good at assigning the highest score to the best translation.

This is where reranking comes in: rescoring with COMET, noisy channel decoding, minimum Bayes risk, etc.

23.01.2025 01:32 β€” πŸ‘ 1    πŸ” 0    πŸ’¬ 1    πŸ“Œ 0

Happy to announce that our work "A Bayesian Optimization Approach to Machine Translation" was accepted to NAACL 2025!

Special thanks to @ufal-cuni.bsky.social for organizing MT Marathon 2025 where I was able to team up with @maikezufle.bsky.social and @zouharvi.bsky.social !

Explainer below:

23.01.2025 01:32 β€” πŸ‘ 11    πŸ” 2    πŸ’¬ 1    πŸ“Œ 1

@juliuscheng is following 19 prominent accounts