#73 - YASAMAN RAZEGHI & Prof. SAMEER SINGH - NLP benchmarks
Patreon: https://www.patreon.com/mlst Discord: https://discord.gg/ESrGqhf5CB YT version: https://youtu.be/RzGaI7vXrkk This week we speak with Yasaman Razeghi and Prof. Sameer Singh from UC Urvine. Yasaman recently published a paper called Impact of Pretraining Term Frequencies on Few-Shot Reasoning where she demonstrated comprehensively that large language models only perform well on reasoning tasks because they memorise the dataset. For the first time she showed the accuracy was linearly correlated to the occurance rate in the training corpus, something which OpenAI should have done in the first place! We also speak with Sameer who has been a pioneering force in the area of machine learning interpretability for many years now, he created LIME with Marco Riberio and also had his hands all over the famous Checklist paper and many others. We also get into the metric obsession in the NLP world and whether metrics are one of the principle reasons why we are failing to make any progress in NLU. [00:00:00] Impact of Pretraining Term Frequencies on Few-Shot Reasoning [00:14:59] Metrics [00:18:55] Definition of reasoning [00:25:12] Metrics (again) [00:28:52] On true believers [00:33:04] Sameers work on model explainability / LIME [00:36:58] Computational irreducability [00:41:07] ML DevOps and Checklist [00:45:58] Future of ML devops [00:49:34] Thinking about future Prof. Sameer Singh https://sameersingh.org/ Yasaman Razeghi https://yasamanrazeghi.com/ References; Impact of Pretraining Term Frequencies on Few-Shot Reasoning [Razeghi et al with Singh] https://arxiv.org/pdf/2202.07206.pdf Beyond Accuracy: Behavioral Testing of NLP Models with CheckList [Riberio et al with Singh] https://arxiv.org/pdf/2005.04118.pdf “Why Should I Trust You?” Explaining the Predictions of Any Classifier (LIME) [Riberio et al with Singh] https://arxiv.org/abs/1602.04938 Tim interviewing LIME Creator Marco Ribeiro in 2019 https://www.youtube.com/watch?v=6aUU-Ob4a8I Tim video on LIME/SHAP on his other channel https://www.youtube.com/watch?v=jhopjN08lTM Our interview with Christoph Molar https://www.youtube.com/watch?v=0LIACHcxpHU Interpretable Machine Learning book @ChristophMolnar https://christophm.github.io/interpretable-ml-book/ Machine Teaching: A New Paradigm for Building Machine Learning Systems [Simard] https://arxiv.org/abs/1707.06742 Whimsical notes on machine teaching https://whimsical.com/machine-teaching-Ntke9EHHSR25yHnsypHnth Gopher paper (Deepmind) https://www.deepmind.com/blog/language-modelling-at-scale-gopher-ethical-considerations-and-retrieval https://arxiv.org/pdf/2112.11446.pdf EleutherAI https://www.eleuther.ai/ https://github.com/kingoflolz/mesh-transformer-jax/ https://pile.eleuther.ai/ A Theory of Universal Artificial Intelligence based on Algorithmic Complexity [Hutter] https://arxiv.org/pdf/cs/0004001.pdf