RT @MilesCranmer: 🧵 Could this be the ImageNet moment for scientific AI?. Today with @PolymathicAI and others we're releasing two massive….

RT @cosmo_shirley: Our internship program at Polymathic is open for opportunities from now through fall 2025! . I believe our program provi….

Join millions who have switched to Grok.

RT @ChatGPTapp: Welcome to the world, OpenAI o1.

RT @cosmo_shirley: You heard all about AI accelerating simulations (maybe from me?), but do you know. How can AI tell you what is in the….

RT @ChanghoonHahn: Excited to announce that our latest #SimBIG collaboration research has just been published in @NatureAstronomy 🔭✨!.#Astr….

Check out our recent work on simulation-based inference in galaxy clustering!.

RT @SiavashGolkar: SOTA models often use bidirectional transformers for non-NLP tasks but did you know causal transformers can outperform t….

RT @MilesCranmer: Some exciting @PolymathicAI news. We're expanding!!. New Research Software Engineer positions opening in Cambridge UK,….

RT @oharub: 🎉 Excited to introduce Gibbs Diffusion (GDiff), a new Bayesian blind denoising method with applications in image denoising and….

RT @YulingYao: We have been using neural posterior/simulation-based inference(SBI) for scientific computing. There was one hole: you run 10….

Blog: Paper: Training Code & Model Weights:

9/ This is joint work with my teammates @PolymathicAI, especially with Francois Lanusse, @SiavashGolkar, @LeopoldoSarra, @MilesCranmer, @cosmo_shirley, and many others!.

8/ The road for these advancements was paved by a lot of exciting work done in SSL for galaxies, especially by @GeorgeStein, @peter_melchior, and @mike_walmsley_.

7/ Ultimately, AstroCLIP’s cross-modal contrastive pre-training produces a high-quality foundation model for galaxies, capable of performing downstream tasks without fine-tuning. We make available the model and code, enabling the community to build higher-level applications.

6/ Morphology classification is another key task for understanding galaxy formation and evolution. AstroCLIP embeddings are informative of morphology and outperform embeddings produced by previous single-modal self-supervised models for galaxies.

5/ AstroCLIP isn't just about redshift estimation. It also effectively embeds information on key galaxy properties like Star Formation Rate, Metallicity, Galaxy Age, and Stellar Mass using simple regression tools. Once again, k-NN outperforms dedicated supervised models.

4/ We evaluated AstroCLIP's performance on photometric redshift estimation, a key inference task in astrophysics. Unlike traditional methods requiring specialized CNNs, AstroCLIP's embeddings are representative enough of the data that even k-NN achieves superior results.

3/ Our updated embedding scheme still aligns galaxy representations across modalities based on shared semantics, providing well-aligned retrievals. Check out below some examples of retrieval, or go to the web app linked above.

2/ The newest AstroCLIP model benefits from a vision transformer backbone trained at scale (300M param.) on the DINOv2 framework, and can now match or outperform dedicated, supervised deep learning models on many downstream tasks without any additional finetuning or training.

1/ AstroCLIP embeds and aligns both galaxy images and optical spectra into a single, physically-meaningful embedding space. AstroCLIP embeddings are sufficiently informative that even simple algorithms like k-NN can be used for a variety of accurate downstream applications.