Specialization of Sub-paths for Adaptive Depth Networks
Lists
Only you see this
Abstract
We present a novel approach to anytime networks that can control network depths instantly at runtime to provide various accuracy-efficiency trade-offs. While controlling the depth of a network is an effective way to obtain actual inference speed-up, previous adaptive depth networks require either additional intermediate classifiers or decision networks, that are challenging to train properly. Unlike previous approaches, our adaptive depth networks require virtually no architectural changes from baseline networks. Instead, we introduce a novel training method that enforces some sub-paths of the baseline networks to have a special property, with which the sub-paths do not change the semantic level of input features, but only refine them to reduce prediction errors. Those specialized sub-paths can be skipped at test time, if needed, to save computation at marginal loss of prediction accuracy. We first formally present the rationale behind the sub-paths specialization, and based on that, we propose a simple and practical training method to specialize sub-paths for adaptive depth networks. While minimal architectural changes and training efforts are required, we demonstrate that our approach significantly outperforms non-adaptive baselines in various tasks, including ImageNet classification, COCO object detection and instance segmentation. Further, we show that the smallest sub-networks of our adaptive depth networks achieve competitive model compression effect compared to recent state-of-the-art techniques.
BibTeX
@inproceedings{kang2023specialization,
author = {Woochul Kang},
title = {{Specialization of Sub-paths for Adaptive Depth Networks}},
booktitle = {International Conference on Learning Representations},
year = {2023},
url = {https://openreview.net/forum?id=yCGgOFC0bG},
}Loading related content...