AI for Science new open source project "Polymathic AI", Yann LeCun serves as consultant

The content that needs to be rewritten is: Editor|Ziluo

In the past few years, by training on a variety of different data sets, visual and natural The field of machine learning for language processing (NLP) has made tremendous progress. This has led to the emergence of "base models" such as "large language models" which have sparked a renaissance in NLP: fine-tuning or prompting generalist models is now standard practice, rather than training specialized models from scratch.

However, the application of machine learning to scientific data sets has yet to undergo a similar paradigm shift.

This is an unrealized opportunity that the “Polymathic AI” (Polymathic AI) research program seeks to address.

Yann LeCun, Turing Award winner and chief scientist of Meta, said: "I am very happy to become a consultant for the new AI for Science project (Polymathic AI)."

Miles Cranmer, assistant professor of AI astronomy/physics at the University of Cambridge, shared on Twitter a new project he is involved in: Polymathic AI!

"We are developing basic models of scientific [data] so that they can take advantage of shared concepts across disciplines."

Netizens expressed: " This is awesome! This looks like fun! This research is amazing..." Basic models for scientific machine learning task customization

The challenge is to build artificial intelligence models that leverage information from heterogeneous data sets and different scientific fields. Contrary to fields such as natural language processing, these models do not share a unified Representation (i.e. text).

These models can then be used as powerful benchmarks or further fine-tuned by scientists for specific applications. This approach has the potential to democratize AI in science by providing ready-made models that are stronger for shared general concepts such as causality, measurement, signal processing, or even more specialized shared concepts such as waves. a priori (i.e. background knowledge). Otherwise, these concepts need to be learned from scratch

To achieve this goal, the research program brings together a team of pure machine learning researchers and domain scientists across various disciplines. In addition, receive guidance from a scientific advisory panel of world-leading experts

research team.

Scientific Advisory Group.

Rewritten to read: Institutional involvement

Building a truly scientifically based model requires a great deal of preliminary research. Our research program is concentrating on the fundamentals of this area. To date, we have published research on key architectural components. Our research spans adapting language models to numerical data, demonstrating the transferability of surrogate models trained on different physical systems, and learning shared embeddings for multimodal scientific data

This research program is important for redefining There is great excitement about the potential of machine learning in science, and Polymathic AI represents an ambitious step towards this goal

Project open source address: https://github.com/PolymathicAI/

Please refer to the following: https://polymathic-ai.org/blog/announcement/

https://polymathic-ai.org/

https://twitter. com/MilesCranmer/status/1711429121220465037

The above is the detailed content of AI for Science new open source project "Polymathic AI", Yann LeCun serves as consultant. For more information, please follow other related articles on the PHP Chinese website!