A Reproducible Benchmark of Lightweight CNNs: Accuracy, Efficiency, and the Impact of Pretrained Initialization
arXiv:2505.03303v3 Announce Type: replace-cross Abstract: Lightweight convolutional neural networks are often compared using results obtained with different training recipes, input settings, and pretrained checkpoints. Such differences make architecture rankings difficult to interpret. This study presents a reproducible benchmark of seven established CNNs across CIFAR-10, CIFAR-100, and Tiny ImageNet under one common fine-tuning protocol. The evaluation reports top-1 accuracy, macro F1, top-5 accuracy, parameter count, FP32 parameter storage, and multiply-accumulate operations. EfficientNetV2-
The proliferation of various lightweight CNN architectures and training methodologies necessitates a standardized benchmark to accurately compare their performance and efficiency.
This benchmark provides critical data for developers and researchers to select appropriate lightweight CNNs for resource-constrained applications, influencing deployment strategies and hardware optimization.
The ability to more reliably compare and select lightweight CNNs based on standardized metrics rather than disparate results becomes possible.
- · AI developers
- · Embedded AI hardware manufacturers
- · Edge computing sector
- · Research institutions
- · Architectures performing poorly in standardized benchmarks
- · Proprietary, non-reproducible benchmark methods
More efficient and accurate deployment of AI models on resource-limited devices will accelerate.
Increased focus on developing architectures that excel in these standardized benchmarks will guide future research and development.
The widespread adoption of highly efficient lightweight CNNs could reduce the overall compute requirements for certain AI applications, potentially impacting demand for high-end GPUs for specific tasks.
This signal links to a primary source. Continuum Brief monitors and indexes it as part of the live intelligence stream — we do not republish source content.
Read at arXiv cs.AI