Our paper got accepted at #AAAI24! 🎉
We propose two new
sequence generation algorithms with "error bars", by adapting beam search to conformal predictions.
If you use LLMs for science, like predicting molecules or proteins that verify some conditions, check it out!
Today we’re launching Proteinbase, a single hub for experimental protein design data. Over 1,000 novel proteins are already live, each with computational predictions, experimental validation, and the method used to design them.
Everything comes from one lab under standardized protocols, which means the results are reproducible, comparable, and include negative data that usually never gets shared.
Zurich AI meetup speakers and friends for 4 Mar!
- Peter Kontschieder, Research Director at Meta. We met back in Mapillary (acq. Meta) days!
- Philippe Schwaller, Asst. Prof at EPFL, AI for chemistry. I love his papers.
- Stef van Grieken, CEO at @cradlebio, AI for proteins. We met back in his Google days!
This'll be a blast :)
Nice write up by Kyle!
I’ll add on - answering questions like "why do we exist" needs no industrial application! It is the *endgame*, sitting atop Maslow's hierarchy. I prefer asking: "how can technology help us understand ourselves and our universe?"
⚡️ New blog post... whilst wearing my Cradle hat!! 🧬
"State-of-the-art enzyme engineering with fully-automated GenAI"
https://t.co/7OOn720z2S
TL;DR: Dataset goes in, SOTA protein function models come out. *Zero human intervention required.*
1/3
Si nous étions en Grande-Bretagne ou aux USA, une telle information ne se serait pas enfouie dans un live mais ferait le gros titre de Une du journal, et une équipe de journalistes serait sur la brèche pour savoir si d’autres candidats ont reçu le même message du Président.
📢 We are hiring! 📢
Are you a MSc/PhD👩🎓👨🎓 passionate about #AI#ML in #cancer, looking for your next career step?
Join our newly-founded AI/ML for Biomedicine group part of the Biomedical Data Science Center at @unil and @CHUVLausanne in beautiful Lausanne!🇨🇭🏔🌈
👇
Interested in Uncertainty Quantification for Sequence Prediction?
Come check out our poster at @RealAAAI today. We add rigorous uncertainty quantification to Beam Search using Conformal Predictions!
@nickgermann@IBMResearch
We have a new preprint on DFT and eigenvalue algorithms:
"Hermitian Pseudospectral Shattering, Cholesky, Hermitian Eigenvalues, and Density Functional Theory in Nearly Matrix Multiplication Time". Link to Arxiv:
https://t.co/zHhVJfy9LS
Announcing torch2jax! Run PyTorch code natively in JAX. 🤝
Mix-and-match PyTorch and JAX code with seamless, end-to-end autodiff, use JAX classics like jit, grad, and vmap on PyTorch code, and run PyTorch models on TPUs.
Excited our Preprint on Conformal Beam Search with @nickgermann and @MariaRoCompBio is out on @Arxiv. We add rigorous uncertainty quantifications to Beam Search using Conformal Predictions. Applicable to any model without changes in the architecture!!
🔗https://t.co/3oE7PRjGNW
One of the cool surprises from this work is how well calibrated the set sizes turned out to be in our experiments with dynamic beam sizes:
Great correlation between how wide the beams need to be vs how they actually are.
Excited to see how well this holds in other exps!
This is relevant for any problem where an autoregressive language model is tasked with generating an exact or approximately exact sequence. Our algorithm generates prediction sets with a guarantee (e.g. 99%) that the correct sequence is present in the set!
Great to see our work @IBMResearch on prompt driven retrosynthesis in @ACSCentSci highlighted by @OPRD_ACS alongside several former colleagues and collaborators!
🌟Highlight: https://t.co/9OZpIPdyQR
📰Paper: https://t.co/ZzDE3CGiOF
#NLP#MachineLearning#Chemistry#AI
Are you dealing with geometric data, be it from molecules or robots? Would you like inductive biases *and* scalability?
Our Geometric Algebra Transformer (GATr 🐊) may be for you.
New work w/ @pimdehaan, Sönke Behrends, and @TacoCohen:
https://t.co/Rvsj1gsKxZ
1/9