Is poisoning ML possible without inserting poison in the model's training set? Yes. @iliaishacked et al. just introduces "data ordering attacks" which are able to target both the integrity and availability of ML simply by *reordering* points during SGDhttps://t.co/4rErkWugiP pic.twitter.com/xx3tNoS64q
— Nicolas Papernot (@NicolasPapernot) April 21, 2021
RoFormer: Enhanced Transformer with Rotary Position Embedding
— AK (@ak92501) April 21, 2021
pdf: https://t.co/2wOHbPx6Ss
abs: https://t.co/LJXnpgFpIK pic.twitter.com/HHCrQ541EV
Facebook is deploying multi-trillion parameter recommendation models into production, and these models are approaching computational intensity of powerful models like BERT. Wrote about research here in Import AI 245: https://t.co/DJpUnl2E78
— Jack Clark (@jackclarkSF) April 20, 2021
Paper here: https://t.co/3LNyUpxavN pic.twitter.com/GpBMmkKApU
Image Super-Resolution via Iterative Refinement
— AK (@ak92501) April 16, 2021
pdf: https://t.co/gzc7bPfYqy
abs: https://t.co/GNRDXh2I6P pic.twitter.com/eViT6nX6L3
(1/2) 🚨 Our new work! 🚨 "Retrieval Augmentation Reduces Hallucination in Conversation" @shtruk @spencerpoff @moyapchen @douwekiela @jasewestonhttps://t.co/r0W6xRuffk
— Jason Weston (@jaseweston) April 16, 2021
We infuse dialogue models with knowledge, significantly reducing hallucinated facts during conversation. pic.twitter.com/1rY1GvYTGg
Curious what we think the future of array computing can look like? Check out our new Dex preprint to see how we’re designing a language for safe, expressive parallelism (incl. no shape errors!) and efficient AD. Available at https://t.co/ae4cOQjoZ8
— Adam Paszke (@apaszke) April 13, 2021
It turns out that vanilla optimizers such as Nesterov momentum and Adam work just as fine for large batch sizes.
— hardmaru (@hardmaru) April 12, 2021
Paper by @zacharynado, @jmgilmer, Chris Shallue, @_arohan_ and George Dahl conducted extensive ablations training vision and language models.https://t.co/KU8pkgAPfn https://t.co/u23KigmXTb
RepVGG: Making VGG-style ConvNets Great Again
— Andrej Karpathy (@karpathy) April 11, 2021
paper: https://t.co/Y5WfgvqxHO
PyTorch code: https://t.co/ydk0RUf6JU
👌Spells out the benefits of very simple/uniform/fast (latency, not FLOPS) deployment architectures. A lot of complexity often due to optimization, not architecture. pic.twitter.com/8GliE4JDiq
New work on Efficient Transformers in RL using Actor-Learner Distillation:
— Russ Salakhutdinov (@rsalakhu) April 10, 2021
Compressing online larger “Learner model” towards a tractable “Actor model” in distributed RL setting with partially-observable environments.https://t.co/jnExWiPabS
with E. Parisotto #ICLR2021 pic.twitter.com/kC5SSRsWrn
InfinityGAN: Towards Infinite-Resolution Image Synthesis
— AK (@ak92501) April 9, 2021
pdf: https://t.co/HrNqpbxnhA
abs: https://t.co/aZX57FbWS5
project page: https://t.co/KhKCzlpC8k pic.twitter.com/tE7UMTsCyE
SiT: Self-supervised vIsion Transformer
— AK (@ak92501) April 9, 2021
pdf: https://t.co/LUmX5fyCbn
abs: https://t.co/ms4ksWdHnD pic.twitter.com/0ALi1YPLTF
Excellent example of low-resource ML research showing AlphaZero scaling laws on a single RTX 2080. https://t.co/MGgwzKJPJq
— Eric Jang 🇺🇸🇹🇼 (@ericjang11) April 8, 2021
