📣 Another announcement for the day!

Next Wednesday (Dec, 14) and Friday (Dec, 16), Luca Manzoni will give a short two-instalments course on The Julia Language: a fresh, innovative programming language especially suited for #ML and #SciComp .

ℹ️ ai2s.it/events/julia-c…

A public service announcement!

(The 'link below' 👉 datascience.sissa.it/event/244/open…)

Interested in #internship or #career opportunities in the fields of #ArtificialIntelligence and other cutting-edge #technologies ?

We are ready to kick-off the social year with the 2nd edition of our #JobFair !

See you there! 👋

🌐 More info: ai2s.it/events/ai2s-jo…

Live in less than 3h! youtube.com/watch?v=edJUp9… (still in 🇮🇹; hopefully without the streaming issues of yesterday)

Live soon, a recommended watch (in 🇮🇹): youtube.com/watch?v=rNEGZ9…

A really great piece by Calvin Luo, organically presenting and demistifying all you need to know about one of the most striking feats of contemporary deep learning. Super recommended! #diffusionmodels

arxiv.org/abs/2208.11970

Today in partnership with EMBL-EBI, we’re releasing predicted structures for nearly all catalogued proteins known to science, which will expand the #AlphaFold database by over 200x - from nearly 1 million to 200+ million structures: dpmd.ai/AF-22-TW 1/

Clever representation or dumb memorization (overfitting)? Our latest #MachineLearning paper models this phase transition, and explains grokking:
arxiv.org/abs/2205.10343

We unify arbitrary vision tasks (here: panoptic, colourisation, monodepth, but could be anything) under a common 'API' that can effectively be learned by a language model. So we get all the benefits of progress in NLP and remove task specific engineering in vision.

The Primacy Bias in Deep Reinforcement Learning

In a new #ICML2022 paper, we identify a damaging tendency of Deep RL agents to overfit to early experiences and propose a simple yet *powerful* remedy by periodically resetting last network layers.

arxiv.org/abs/2205.07802
1/N 🧵

why do we even want 'agi' if the way of achieving it will be as boring as scaling up a neural net and sgding it on the internet?

i mean, the only reason i find working towards agi even mildly interesting is the chance of learning something about cognition in the process...