Quantum entanglement for attention models
Keywords: Attention models, Quantum entanglement, Transformers
TL;DR: We study quantum entanglement entropy in attention models.
Abstract: Attention mechanisms in deep learning establish relationships between different positions within a sequence, enabling models like Transformers to generate effective outputs by focusing on relevant input segments and their relations. The performance of Transformers is highly dependent on the chosen attention mechanism, with various approaches balancing trade-offs between computational cost, memory efficiency, and generalization ability based on the task.
Quantum machine learning models possess the potential to outperform their classical counterparts in specialized settings. This makes exploring the benefits of quantum resources within classical machine learning models a promising research direction. The role of entanglement in quantum machine learning, whether in fully quantum or as subroutines in classical-quantum hybrid models, remains poorly understood. In this work, we investigate whether quantum entanglement, when used as a resource, can improve the performance of the attention layer in Transformers.
We introduce an entanglement-based attention layer within a classical Transformer architecture and numerically identify scenarios where this hybrid approach proves advantageous. Our experiments on simple standard classification tasks in both vision and NLP domains reveal that the entanglement-based attention layer outperforms classical attention, showing superior generalization on quantum-generated datasets and in settings with limited training data for classical datasets. Additionally, it demonstrates a smaller generalization gap across all tested datasets. Our work contributes towards exploring the power of quantum resources as a subroutine in the classical-quantum hybrid setting to further enhance classical models.
Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 10491
Loading