(5/5) Perhaps I'm overlooking something-open to your insights 😀

(4/5) However, what we see in DeepSeek's formulation is that 1) there is no sequence of actions, and 2) the reference policy remains the same.

I don't see the credit assignment mechanism from future rewards to current actions in this formulation, which is the key factor in RL.

(3/5) Because this objective is difficult to compute (there are two distributions from two different policies involved) a constrain on how far these two policies can be is included so we can just "ignore" this mismatch.

(2/5) In TRPO and PPO, you maximize the expected sum of advantages over a sequence of actions as a way to optimize policy improvement steps.

I still uncertain about the RL aspect in DeepSeek.

To me it looks like a clever way of applaying a PPO-like clipping within a supervised framework, constrained by a fixed reference model. Althought some parts in its formulation are very similar to PPO, I wouldn't describe it as RL. (1/5)🧵

Any recommendation for a drug discovery related confernece in Europe this year? Im all ears 😄

*generalize

softmax is not enough (for sharp out-of-distribution) A key property of reasoning systems is the ability to make sharp decisions on their input data. For contemporary AI systems, a key carrier of sharp behaviour is the softmax function, with its capabili...

Why your transformer-based net does not generalized from small molecules to peptides? Well, here one of the reasons: arxiv.org/abs/2410.01104

Once you know the root of the problem, you can find nice solutions 😉
In our case, a very simple regularization term did the job.

Great book. I also enjoyed it a lot.

Thanks! We introduce inductive bias at different levels. For instance, we refined the centrality encoder to implicitly capture atom hybridization. We have other examples in the paper, and many others that we hope to publish soon 😉

I'm making a list of AI for Science researchers on bluesky — let me know if I missed you / if you'd like to join!

go.bsky.app/AcP9Lix

Same here! Thanks :)

Thanks a lot!

Analysis of Atom-level pretraining with Quantum Mechanics (QM) data for Graph Neural Networks Molecular property models Despite the rapid and significant advancements in deep learning for Quantitative Structure-Activity Relationship (QSAR) models, the challenge of learning robust molecular representations that effectiv...

I'm working on improving molecular GNN models by adding inductive bias and integrating different data modalities. We recently presented part of our work in a workshop at ICML last July: arxiv.org/abs/2405.14837

Thanks a lot!

A starter pack for anyone interested in AI & drug discovery :)

go.bsky.app/AgYHc8j

Hi, just arrived here. Nice to see this starter pack. Im also doing some research in drug design and I would love to be inluded there as well. Thanks!

Follow leading researchers, practitioners, and thought leaders exploring the intersection of #AI, machine learning, data science, and scientific research. 🦋

➕We are just getting started, please send me others who should be added

go.bsky.app/JeFdryY #ML

Great initiative! I'd love to be there as well