Junsong Yuan

Congratulations to Qianlin Liang on successfully defending his Ph.D thesis titled "Resource Management for Edge AI." Qianlin's thesis looks at model-driven resource management for edge AI workloads. #edgeAI #cloudcomputing

93

7 Comments

Interested in solving combinatorial #scheduling problems with differentiable models and gradient descent on GPUs, TPUs, or other accelerators? Coming at ICML 2024 #ICML, Mingju Liu (Mingju's website: https://lnkd.in/gbyWAwT9 ) present our paper, "Differentiable Combinatorial Scheduling at Scale," where we introduce a fully differentiable modeling approach for combinatorial scheduling. This new formulation allows for a fully differentiable encoding of SDC-scheduling #sdc #scheduling, which can be extended to broader combinatorial problems under linear programming (LP) models. Great thanks to the scheduling/HLS expert Prof. Zhiru Zhang from Cornell for many insightful discussions and advice, and congrats to our leading student authors Mingju Liu Yingjie Li. If you are interested in solving discrete combinatorial problems with differentiable models using ML infrastructures, please check out our paper and GitHub repository, and feel free to reach out! Paper: https://lnkd.in/gFCyGXZX Code: https://lnkd.in/gUik3NuA #ICML2024 #Scheduling #Opensource

125

2 Comments

This was a fun paper to work on, and many thanks to Jacy Reese Anthis and Kristian Lum for leading it and inviting me to join. Thought-provoking is exactly what we were going for, Xavi! Want to quick take a moment to provide a bit of my own background story for this work, because I don't think I've actually posted much on it yet. The perspective I brought to this collaboration comes from my attempts to synthesize Selbst et al.'s arguments about abstraction (https://lnkd.in/e88iaznN), Dwork and Ilvento on composition (https://lnkd.in/e5_9-zyn), our findings trying to map out the problem space of fair IR and recsys (https://lnkd.in/gF4v8vcY, https://lnkd.in/edn_ecjz), and other things I've been reading, reaching the conclusion that effective fair (and, more generally, responsible) AI/info system work needs to be grounded in specific harms or goals, and that making them "fair" in a universal sense is impossible due to the nature of fairness as a social, essentially-contested construct. I believe we can derive general knowledge from this effort inductively, identifying principles and methods that can be applied across multiple settings, but the problems — and therefore the evaluation and specific mitigation techniques — are deeply contextual and specific. This also does not mean that we should not study or seek to improve various forms of fairness in models intended to power a wide range of applications, but that doing so will not produce a universally "fair" model and is insufficient to guarantee fairness of downstream applications. Hope many of you also appreciate this paper!

31

2 Comments

Amid the excitement of two Nobel prizes awarded to AI leaders this year, my first GPT paper is out! (No, we did not write it with GPT... or did we? 🤔 Who can tell these days?) All jokes aside, my colleagues Xin Wang, Liangliang Huang, Kun Lu, and I explored the ability of GPT-4 to extract band gap energies from 415 abstracts. With the right prompt engineering, GPT-4 achieved an impressive 94% accuracy, a huge leap from the 51% accuracy we saw using a traditional rule-based method. Check it out and let us know what you think! #AI #GPT4 #MachineLearning #InformationExtraction #BandGap

26