AI System Helps Scientists Write Expert-Level Software
▼ Summary
– ERA is an AI system that uses a Large Language Model and Tree Search to create expert-level scientific software aimed at maximizing a quality metric.
– In bioinformatics, ERA discovered 40 novel single-cell data analysis methods that outperformed top human-developed methods on a public leaderboard.
– In epidemiology, ERA generated 14 COVID-19 hospitalization forecasting models that outperformed the CDC ensemble and all other individual models.
– ERA also produced expert-level software for geospatial analysis, neural activity prediction, numerical integration, and a novel time series forecasting rule-based construction.
– By devising and implementing novel solutions across diverse tasks, ERA represents a significant step towards accelerating scientific progress.
Scientific breakthroughs are often slowed by the painstaking, manual development of software needed to run computational experiments. To break this logjam, researchers have introduced the Empirical Research Assistance (ERA) system, an AI designed to produce expert-level scientific software with a singular focus: maximizing a specific quality metric. At its core, ERA combines a Large Language Model (LLM) with Tree Search (TS) algorithms, allowing it to systematically refine that metric while intelligently exploring a vast array of potential solutions.
The system truly shines when it ventures beyond its training data, actively seeking out and integrating complex research concepts from external sources. This tree search capability proves its worth across a wide spectrum of scientific challenges. In bioinformatics, ERA unearthed 40 novel methods for single-cell data analysis, all of which surpassed the best human-developed techniques on a public leaderboard. In epidemiology, it generated 14 forecasting models that outperformed both the CDC ensemble and every other individual model in predicting COVID-19 hospitalizations.
ERA’s reach extends even further. It has produced expert-level software for geospatial analysis, predicting neural activity in zebrafish, and numerically solving integrals. It also crafted a novel rule-based construction for time series forecasting. By devising and implementing original solutions to such diverse tasks, ERA marks a substantial leap forward in accelerating the pace of scientific discovery.
(Source: Nature.com)
