Daily Arxiv

This is a page that curates AI-related papers published worldwide.
All content here is summarized using Google Gemini and operated on a non-profit basis.
Copyright for each paper belongs to the authors and their institutions; please make sure to credit the source when sharing.

CRINN: Contrastive Reinforcement Learning for Approximate Nearest Neighbor Search

Created by
  • Haebom

Author

Xiaoya Li, Xiaofei Sun, Albert Wang, Chris Shum, Jiwei Li

Outline

CRINN presents a new paradigm for optimizing the approximate nearest neighbor search (ANNS) algorithm using reinforcement learning. By treating ANNS optimization as a reinforcement learning problem with execution speed as a reward signal, CRINN automatically generates progressively faster ANNS implementations while maintaining accuracy constraints. In experimental evaluations on six popular ANNS benchmark datasets, CRINN achieves state-of-the-art performance on three datasets (GIST-960-Euclidean, MNIST-784-Euclidean, and GloVe-25-angular) and first place on two datasets (SIFT-128-Euclidean and GloVe-25-angular) compared to state-of-the-art open-source ANNS algorithms. We demonstrate that reinforcement learning-augmented LLMs can serve as an effective tool for automating sophisticated algorithmic optimizations that previously required expert knowledge and extensive manual effort. The code can be found at https://github.com/deepreinforce-ai/CRINN .

Takeaways, Limitations

•
Takeaways:
◦
Presenting new possibilities for optimizing the ANNS algorithm using reinforcement learning.
◦
Demonstrating the effectiveness of automated algorithm optimization through the combination of LLMs and reinforcement learning.
◦
Proven practicality with excellent performance on multiple benchmark datasets
◦
Suggesting the possibility of automating the algorithm optimization process, which requires specialized knowledge and manual work.
•
Limitations:
◦
Generalization performance verification on datasets other than the presented benchmark dataset is needed.
◦
Analysis of the computational cost and training time of reinforcement learning-based optimization processes is needed.
◦
Further research is needed on the applicability and generalization performance of various ANNS algorithms and problem settings.
👍
Made with Slashpage