weekend yt lecture suggestion:
“What's wrong with #LLMs
&
what we should be building instead”
ಠ_ಠ
issues:
• Incorrect & contradictory answers
• Dangerous & socially unacceptable answers
• Expensive to train & lack of updatability
• Lack of attribution & poor non-linguistic knowledge
challenges:
• Sepparate Linguistics from World Knowdledge
• Functioning Episodic Memory
By Veteran since #GOFAI times, #MLPioneer #JMLR Journo etc:
Prof #ThomasGDietterich @tdietterich https://youtu.be/cEyHsMzbZBs
This paper considers the problem of choosing "good" #MachineLearning algorithms from their performance on a number of datasets. They use an idea from #psychometric testing called "item-response theory". A particular benchmark is a single question in a test (some datasets are hard, others are easy). A particular ML algorithm is a student taking a test (some algorithms are consistent, some do well on easy questions, others give puzzling performance).
'Sufficient reductions in regression with mixed predictors', by Efstathia Bura, Liliana Forzani, Rodrigo Garcia Arancibia, Pamela Llop, Diego Tomassi.
'Towards An Efficient Approach for the Nonconvex lp Ball Projection: Algorithm and Analysis', by Xiangyu Yang, Jiashan Wang, Hao Wang.
'Total Stability of SVMs and Localized SVMs', by Hannes Köhler, Andreas Christmann.
'Distributed Learning of Finite Gaussian Mixtures', by Qiong Zhang, Jiahua Chen.
'PECOS: Prediction for Enormous and Correlated Output Spaces', by Hsiang-Fu Yu, Kai Zhong, Jiong Zhang, Wei-Cheng Chang, Inderjit S. Dhillon.
'Unlabeled Data Help in Graph-Based Semi-Supervised Learning: A Bayesian Nonparametrics Perspective', by Daniel Sanz-Alonso, Ruiyi Yang.
'Rethinking Nonlinear Instrumental Variable Models through Prediction Validity', by Chunxiao Li, Cynthia Rudin, Tyler H. McCormick.
'Attraction-Repulsion Spectrum in Neighbor Embeddings', by Jan Niklas Böhm, Philipp Berens, Dmitry Kobak.
Teixi
Cheng Soon Ong
JMLR