New machine learning research article from Atinary

Atinary announces their latest research on Multi-fidelity Optimization (MFBO), a machine learning technique that optimizes several information sources with different fidelities.

Posted
October 17, 2024
Author
Edlyn W
Category
Share this article

How to balance accuracy and cost in experiments? 🎯⚖️💲

We’re excited to announce Atinary’s latest research on Multi-fidelity Optimization (MFBO) that tackles this R&D challenge, in collaboration with Profs. Jose Miguel Hernández Lobato (University of Cambridge), Jeremy Luterbacher (EPFL) and Philippe Schwaller (EPFL) and their teams. 

Scientists often face a tough choice: 🤔
➡️ Run an accurate but expensive experiment? (High-fidelity); or
➡️ Opt for a less precise yet more affordable alternative? (Low-fidelity)

Our team has developed best practices for MFBO, a machine learning technique that optimizes several information sources with different fidelities.

Our research showcases:
📈 Specific experimental settings where combining high- and low-fidelity experiments outperforms traditional (or single-fidelity) BO approaches.
💰 Optimal cost ratios (cost for low-fidelity over high-fidelity experiment) for effective resource allocation.
👩‍🔬 Successful use of MFBO to real life applications in molecular and materials optimization achieving up to a 70% reduction in the number of experiments. 

This research offers scientists a roadmap by:
✅ Reducing costs by up to 70% in optimization by leveraging cheap information sources.
✅ Accelerating optimization by exploiting complementary and cheap sources of information in materials and molecular research to attain cost efficiency.
✅ Providing best practices on when to use MFBO over single-fidelity BO.

👏 Atinary is committed to advancing AI-driven scientific innovation. Congrats to all authors on this exciting collaborative research: Victor Sabanza Gil, Daniel Pacheco Gutierrez, Jeremy S. Luterbacher, Riccardo Barbano, José M. Hernández-Lobato, Philippe Schwaller, and Loïc Roch

🗞️ For more details, read the full article on arXiv: https://arxiv.org/html/2410.00544v1