ML CoPilot, a collaborative approach of large language models (LLMs) and human expertise, proposes the use of vector databases to store and retrieve past machine learning (ML) insights for suggesting solutions to new ML tasks. The authors suggest ML CoPilot can learn from past ML benchmarks, HPO-B, PD1, and HyperFD, which represent a wide range of ML tasks, to form a knowledge repository in a text format, comprehensible by LLMs. When presented with a new ML task, ML CoPilot searches for related tasks in the pool and provides the gathered information to LLMs, which then generate the solution for the given problem, resulting in more accurate and efficient solutions to novel ML tasks.

source update: Empowering Large Language Models with Human… – Towards AI