Binarized Convolutional Neural Networks with Channel Quadrupling and Smooth Downsampling
Keywords: Binarized Convolutional Neural Networks, Image Classification, Image Segmentation, Mobile-Friendly Convolutional Neural Networks
TL;DR: his paper proposes a novel BCNN model designed to quadruple the number of channels and incorporate a so-called smooth downsampling in BCNNs for mobile environments.
Abstract: This paper proposes novel binarized convolutional neural networks (BCNNs) named **QB-Net** and **QSB-Net**, specifically designed to **Q**uadruple the number of channels and incorporate a so-called **S**mooth downsampling in **B**CNNs for low-cost mobile environments. The proposed models combine FP32 depthwise separable (DS) convolutions with binarized $1 \times 1$ pointwise convolutions, offering reduced computational costs in the pointwise convolutions. To enhance the degraded performance of the above naive combination, the proposed models start with a small number of channels in shallow layers and expand them during downsampling by a factor of four, effectively managing model complexity in the downsampling. The proposed model structure maintains low computational costs in the shallow blocks and increases model complexity in the deep blocks, providing a wider dynamic range to manage information in the frequency domain. As a result, the proposed models overcome the limitations of existing BCNNs, delivering improved performance while reducing the total computational costs. For further performance enhancements, we propose a novel smooth downsampling with heightwise and widthwise sequential downsampling steps, doubling the number of channels at each step. Besides, we show that the channelwise self-attention (SE) is applicable with minimal additional computational costs in the proposed models. Besides, multiple binarized convolutions in the fully-connected (FC) layer reduce storage costs without requiring 8-bit quantized convolutions. Experimental results demonstrate the efficiency of the proposed models in terms of performance, computational costs, and inference latency on real hardware. Notably, the QSB-Net-Large with SE achieve 71.2\% Top-1 accuracy on ImageNet-1K and 69.2 mean intersection over union (mIoU) in the semantic segmentation on the PASCAL VOC dataset, outperforming other counterparts.
Supplementary Material: zip
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: 5505
Loading