Neural ODE Transformers: Analyzing Internal Dynamics and Adaptive Fine-tuning
Keywords: neural ODEs, transformer, adaptive finetune
TL;DR: This paper introduces a new transformer architecture using non-autonomous neural ODEs that allows adaptive fine-tune. The paper analyze internal model dynamic using spectral analysis.
Abstract: Recent advancements in large language models (LLMs) based on transformer architectures have sparked significant interest in understanding their inner workings. In this paper, we introduce a novel approach to modeling transformer architectures using highly flexible non-autonomous neural ordinary differential equations (ODEs). Our proposed model fully parameterizes all the weights of attention and feed-forward blocks through neural networks, with weights articulated as functions of a continuous layer index. We examine the model's dynamics through spectral analysis, uncovering an increase in eigenvalue magnitude offering a practical insights against weight-sharing assumption in existing theoretical studies. We also introduce the use of the Lyapunov exponent to examine token-level sensitivity, improving model interpretability. Our neural ODE transformer performs similarly to GPT across various configurations and datasets, while offering flexible fine-tuning capabilities under different architectures.
Supplementary Material: zip
Primary Area: foundation or frontier models, including LLMs
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.
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: 3683
Loading