NetBooster: Empowering Tiny Deep Learning By Standing on the Shoulders of Deep Giants

Zhongzhi Yu, Yonggan Fu, Jiayi Yuan, Haoran You, Yingyan Lin

Keywords: Network Training, Transfer Learning

TL;DR: We propose an expansion-then-contraction training strategy on both width and depth dimension to fully unleash tiny neural network's potential on large scale datasets and downstream tasks.

Abstract: Tiny deep learning has attracted increasingly growing interest driven by the substantial demand for deep learning solutions in numerous Internet-of-Things (IoT) applications. Nevertheless, due to the under-fitting issue, it is still a challenge to unleash tiny deep learning’s full potential on large-scale datasets. Consequently, tiny neural networks’ (TNNs’) downstream task performance is limited due to the inferior learned representations during pretraining. To this end, we propose a framework dubbed NetBooster which empowers tiny deep learning from a novel perspective by augmenting the architecture of TNNs via an expansion-then-contraction strategy. Specifically, during training, our proposed NetBooster first expands each/some layer(s) of a given TNN into multi-layer blocks, favoring the learning of more complex features to generate an expanded counterpart model (i.e., deep giant), and then contracts the expanded layers by gradually removing the non-linear layers from the expanded ones to recover efficiency. NetBooster’s expansion-then-contraction training empowers its trained TNNs to benefit from the superior performance of their expanded counterparts while preserving the TNNs’ original complexity and thus inference efficiency. Extensive experiments and ablation studies on two tasks, seven datasets, and six networks validate that NetBooster consistently leads to a nontrivial accuracy boost (e.g., 1.3% ∼ 2.5%) on top of state-of-the-art TNNs on ImageNet and as much as 4.7% higher accuracy on various downstream datasets, while maintaining their inference complexity/efficiency.

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.

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: Yes

Please Choose The Closest Area That Your Submission Falls Into: Deep Learning and representational learning