Stanford AI Lab

We've tried SO many fancy multi-task optimizers published in top ML conf. venues for "simple" regulatory genomics tasks and so far none of them can beat the simple: training naive multi-task model -> fine tune on each task. https://twitter.com/y0b1byte/status/1481283351281553417

yobibyte @y0b1byte

🚨 Paper Alert https://arxiv.org/abs/2201.04122
The multitask learning 📔📜 claims that naive grad summation is weaker than ad-hoc optimizers due to task interference/grad conflicts. We show that you can get the same results by coupling it with plain old techniques, e.g. regularisation 🧵

Work already starting for this year's conference #NeurIPS2022. Please nominate yourself to give us the widest set of people to bring into the organising roles this year. Excited to be a co-General Chair with @sanmikoyejo 🥳 https://twitter.com/neuripsconf/status/1486290789768564740

Excited to share our #iclr2022 paper "GreaseLM: Graph REASoning Enhanced Language Models for Question Answering", w/ the amazing @ABosselut @michiyasunaga @ren_hongyu @percyliang @chrmanning and @jure!

Paper: https://arxiv.org/abs/2201.08860
Code: https://github.com/snap-stanford/GreaseLM

🧵👇[1/7]

Twitter feed image.

@StanfordAILab Very cool work just posted on the @StanfordAILab blog! An approach for training robotics controllers by yelling at it instead of painstakingly annotating each frame.