Bill Byrne

@mmbronstein @desirivanova @PetarV_93 @chaitjo @fhuszar @inxoy_ @bose_joey @pl219_Cambridge I've given lectures and practical exercises on transformers as part of the Cambridge MLMI MPhil since at least 2021. The module focused on machine translation, dialogue, and other NLP.

@desirivanova @PetarV_93 @chaitjo @fhuszar @inxoy_ @bose_joey @pl219_Cambridge We teach what transformers are

Not sure about Cambridge, but Oxford courses are like ~10 years out of date. In stats, we are still teaching SVMs and kernels in the "advanced" ML classes. AFAICT, there aren't any serious DL classes in other departments (CS, math) either, of the likes of e.g. Stanford's offering

@ianwu97 arxiv.org/abs/2311.08380

@ianwu97 Nice to see this approach applied to tasks other than NMT. arxiv.org/pdf/2410.02902

1/ What’s the best way to supervise LLMs using LLM judges? We show that Minimum Bayes Risk (MBR) decoding is the way to go! With MBR decoding you can: 💡 Trade compute for performance at inference time 🧰 Generate data for self-training without needing external labels

Time to dust off all the old Boltzmann Machine papers...

@kchonyc ok, that's funny .

it's taking surprisingly long for an LLM bro to announce the discovery of task oriented dialogue systems

@natolambert Very nice work. Following your notes in the discussion, it'll be really interesting to see if the `reference free` reward for DPO also works. Having the reward depend on the regularisation mechanism used in training doesn't seem quite right.

Excited to share something that we've needed since the early open RLHF days: RewardBench, the first benchmark for reward models. 1. We evaluated 30+ of the currently available RMs (w/ DPO too). 2. We created new datasets covering chat, safety, code, math, etc. We learned a lot. We hope this is a major step in understanding why reward models work, rather than just how they do for RLHF. In short, we created pairs of responses, one good one bad (with manual review) and see where reward models agree! It's a simple and powerful process. Key takeaways: * Running reward models is hard, we build infra to make this easier. * We're already using this to learn more about PPO RLHF training (more on this soon). * Reward models mirror the refusals behavior we're confused about in RLHF. Some refuse everything (including llama 2 style stuff), some refuse nothing, and few models handle both cases well. * Datasets like Anthropic HH / Learning to Summarize only take us so far (and don't work for DPO) * Scaling matters (big models win again) Here's the current leaderboard: I'm very excited about future work. Figuring out what values are reflected, generative RMs, better RMs for training, more on DPO, and everything in between. Links! Leaderboard: huggingface.co/spaces/allenai… Code: github.com/allenai/reward… Paper (arxiv soon): github.com/allenai/reward… Eval dataset: huggingface.co/datasets/allen… YouTube walkthrough: youtu.be/CAaHAfCqrBA

@JChiyahGarcia @ale_suglia @arash_eshghi @hfhastie And well done @alexcoca23, Bo-Hsiang Tseng, Jinghong Chen, Weizhe Lin, Weixuan Zhang, Tisha Anders, Bill Byrne for winning best long paper!! (Let me know if we have failed to find your handle and you want to be tagged)

When you register for @EAMT_2023 , don't forget to sign up for our workshop (it's free!) so you can hear our 2nd 🔑🎵 Danielle Saunders talk about the challenges and needs in #Gender #inclusive #MachineTranslation, such as applying methods at scale (data size & language variety)

Check out our survey on dense retrieval & QA for low-resource scenarios that was accepted at #EACL2023 @GianniBarlacchi @marcodeltredici @unattributed @AmazonScience #AmazonScience amazon.science/publications/n…

@bnjmn_marie although there may have been earlier internal work at IBM. But here's a cite for MBR system combination ieeexplore.ieee.org/abstract/docum…

Since MBR decoding is making a comeback, I gave it another try for MT system combination (so not during decoding itself as in recent work) using COMET. medium.com/mlearning-ai/m…

@bnjmn_marie Nice writeup! Its great to see early work by Vaibhava Goel cited, but that 2000 paper was about A* search over lattices. Andreas Stolcke and colleagues had a'97 paper (MBR over n-best lists), which we cite, and which is probably the earliest citation in ASR...

@zngu i see - thanks!

@zngu Agree entirely about UK visa process. Do you have details on the Huawei scheme? How does it avoid the 'working away' problem?

@margs_li Looking forward to reading. For some prior work on Bayesian domain weighting in NMT, check out aclanthology.org/P19-1022/

Train an LM made of independent expert LMs (no syncs! no shared params!) ➡️ ➕ new or ➖ existing experts. At. Any. Time. ➡️ Ensemble OR parameter average(!!) to outperform dense & sparse LMs & ensemble baselines with less compute, a fraction of the simultaneous GPU usage. 🌳/n

Interested in doing a 4-6 month research internship on topics relevant to #convsearch or QA? Our team at Amazon Alexa AI is looking for interns at Barcelona & Berlin locations! Reach out directly and apply via amazon.jobs/en/jobs/173291… @chengweiwei @GianniBarlacchi @marcodeltredici

Adapting machine translation models to new genres amazon.science/blog/adapting-…

Tomorrow we will be hosting our free event on "Understanding and Automating #Counterspeech". We are very excited to discuss this important topic with experts on philosophy of language, law, computer science, media & political communication, counter terrorism and online activism.

@random_walker and another paper on coaxing map to avoid bias - arxiv.org/abs/2104.07429

@random_walker Here's a nice recent paper on avoiding map to avoid bias - arxiv.org/abs/2105.08504

When a machine learning system uses argmax to select outputs from a probability distribution — and most of them do — it's a clue that it might be biased. That's because argmax selects the "most probable" output, which may amplify tiny data biases into perfectly biased outputs.