Adaptive Inference: Theoretical Limits and Opportunities for Efficient AI

Soheil Hor, Ying Qian, Mert Pilanci, Amin Arbabian

Keywords: Adaptive Inference, Efficient ML, Dynamic Neural Networks, Dynamic Routing, Computer Vision, Natural Language processing

TL;DR: A new theoretical framework for quantifying efficiency and performance gains achievable through adaptive inference

Abstract:

With the commercial deployment of increasingly larger and more complex neural networks at the cloud and the edge in recent years, inference has become too costly in terms of compute workload worldwide. Adaptive inference methods, which dynamically adjust a neural network's size or structure during inference, offer a means to enhance efficiency of neural networks beyond what static network compression and optimization methods can fundamentally achieve.

This paper introduces the first theoretical framework for quantifying the efficiency and performance gain opportunity size of adaptive inference algorithms. We provide new approximate and exact bounds for the achievable efficiency and performance gains, supported by empirical evidence demonstrating the potential for 10-100x efficiency improvements in both Computer Vision and Natural Language Processing tasks without incurring any performance penalties. Additionally, we offer insights on improving achievable efficiency gains through the optimal selection and design of adaptive inference state spaces.

Primary Area: other topics in machine learning (i.e., none of the above)

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: 8153