Kernel Autoresearch discovers reusable inductive biases for ML models

Researchers have developed Kernel Autoresearch (Kernaut), a novel approach to discovering new machine learning kernels. Kernaut treats kernel design as an open-ended model discovery process, where coding agents generate kernels as programs and construction contracts ensure their validity. This method utilizes a quality-diversity archive to store high-performing kernels with distinct behaviors and employs novelty screening to guide agents toward functionally new candidates. Experiments show that Kernaut-discovered kernels encode reusable inductive biases that generalize to unseen tasks, outperforming meta-learned deep kernels and traditional baselines on various optimization and enzyme-kinetic rate law problems. Furthermore, these discovered kernels are interpretable programs that can be refined by human researchers to further improve performance. AI

IMPACT Introduces a novel method for generating interpretable and reusable machine learning kernels, potentially improving model generalization and research efficiency.

RANK_REASON The item is an arXiv preprint detailing a new research methodology for machine learning kernel discovery. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• alphaXiv
• arXiv
• CatalyzeX
• Connected Papers
• DagsHub
• Gotit.pub
• Hugging Face
• Kernaut
• Kernel Autoresearch
• Litmaps
• Richard Cornelius SUWANDI
• ScienceCast
• scite Smart Citations

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →