François Chollet: The worst-performing case also deter...| Robocrunch AI

The worst-performing case also determines deployability for most high-value systems.
— François Chollet (@fchollet) May 12, 2022

Related

Similarly, if ablating a fancy component kills performance, it's worth investigating whether it's not just because you have fewer parameters ove ...

Shared by Zenna Tavares at 4/18/2022 0-3

How to avoid system crashing when trying a new large model: 1. run systemd-run --user --scope -p MemoryMax=100G -p MemorySwapMax=50G bash 2. ...

Shared by Stas Bekman at 5/20/2022 0-6

https://huyenchip.com/2021/09/13/data-science-infrastructure.html

1/ A problem everyone faces in machine learning is getting a model that you've trained deployed for production-level traffic. Hard because dev ...

Shared by Abubakar Abid at 4/1/2022 1-4

https://twitter.com/bitman_mars/status/1514269034094764046

Someone should try to inspect the model output conditioned on those high loss batches just to make sure it does not have structure. That it does ...

Shared by Andrej Karpathy at 4/13/2022 4-71

I'm no hardware expert, but - if you need 2x the power and (potentially) 2x the price to 3x the compute - it seems to me that hardware has littl ...

Shared by Jonathan Frankle at 3/23/2022 0-12

https://arxiv.org/abs/1911.05248

This was one of my hopes in both https://t.co/5yg2p1zlG2 and https://t.co/lXWsnPyEza -- that differences in capacity provide a powerful auditing ...

Shared by Sara Hooker at 11/18/2021 2-7

768