GroupNL: Low-Resource and Robust CNN Design over Cloud and Device

TL;DR

This work tackles robust, efficient CNN deployment for cloud-assisted IoT by eliminating reliance on cheap convolutions and BN in favor of grouped seed features and fixed nonlinear transformations. The proposed GroupNL design generates diverse feature maps with $|\ Psi_{gnl}| = \gamma g$ nonlinearities while keeping parameter and FLOP counts low, and includes a sparse variant to further reduce cost. Empirical results show GroupNL improves accuracy on Icons-50 and ImageNet-C, accelerates multi-GPU training (e.g., up to ~11% speedup on ResNet-101 and 53% on ImageNet-1K with 8 GPUs), and delivers favorable on-device performance on a Raspberry Pi, with notable reductions in model size for large networks. Overall, GroupNL offers a practical, robust, low-resource CNN design for cloud-device deployment, with promising directions for theoretical analysis of NLF choices and more targeted ablations.

Abstract

Deploying Convolutional Neural Network (CNN) models on ubiquitous Internet of Things (IoT) devices in a cloud-assisted manner to provide users with a variety of high-quality services has become mainstream. Most existing studies speed up model cloud training/on-device inference by reducing the number of convolution (Conv) parameters and floating-point operations (FLOPs). However, they usually employ two or more lightweight operations (e.g., depthwise Conv, $1\times1$ cheap Conv) to replace a Conv, which can still affect the model's speedup even with fewer parameters and FLOPs. To this end, we propose the Grouped NonLinear transformation generation method (GroupNL), leveraging data-agnostic, hyperparameters-fixed, and lightweight Nonlinear Transformation Functions (NLFs) to generate diversified feature maps on demand via grouping, thereby reducing resource consumption while improving the robustness of CNNs. First, in a GroupNL Conv layer, a small set of feature maps, i.e., seed feature maps, are generated based on the seed Conv operation. Then, we split seed feature maps into several groups, each with a set of different NLFs, to generate the required number of diversified feature maps with tensor manipulation operators and nonlinear processing in a lightweight manner without additional Conv operations. We further introduce a sparse GroupNL Conv to speed up by reasonably designing the seed Conv groups between the number of input channels and seed feature maps. Experiments conducted on benchmarks and on-device resource measurements demonstrate that the GroupNL Conv is an impressive alternative to Conv layers in baseline models. Specifically, on Icons-50 dataset, the accuracy of GroupNL-ResNet-18 is 2.86% higher than ResNet-18; on ImageNet-C dataset, the accuracy of GroupNL-EfficientNet-ES achieves about 1.1% higher than EfficientNet-ES.

Figures (13)

• Figure 1: Overview of cloud-assisted architecture.
• Figure 2: The example of GroupNL Conv with $c_\mathrm{in} \!=\! 8, c_\mathrm{out}\!=\! 16$, and the number of seed features is $c_\mathrm{seed} \!=\! 4$, the groups for splitting is $g\!=\!2$. In this case, the GroupNL can introduce $\gamma \cdot g \!=\! 6$ different NLFs for the feature maps generation. Compared with the vanilla Conv, which needs 16 filters to compute all features, the GroupNL Conv only requires 4 filters to compute the seed features and generates the remaining features by nonlinearities and tensor manipulation operators.
• Figure 3: The visualization of feature maps generated based on the diverse NLFs with 24 different hyperparameter settings. The diversity of the generated feature maps is greatly improved while maintaining the similarity.
• Figure 4: Visualization examples of different severity levels of corruptions on ImageNet-C.
• Figure 5: Hardware: a Resource-constrained Raspberry Pi 4B for Deployment and a Power Meter for Energy Evaluation.