Adaptive EfficientNet: Dynamic Compound Scaling for Resource-Optimal Deep Learning Article Swipe
YOU?
·
· 2025
· Open Access
·
· DOI: https://doi.org/10.5281/zenodo.17822377
The widespread adoption of deep learning models in diverse applications often clashes with the stringent resource constraints of deployment environments, particularly for mobile and edge devices. Traditional deep neural networks, while achieving remarkable accuracy, typically demand significant computational power, memory, and energy. EfficientNet architectures introduced a systematic compound scaling method to efficiently scale model width, depth, and resolution, yielding superior performance across various models. However, this scaling is predominantly static, pre-determined during design, and lacks adaptability to fluctuating or dynamic resource availability. This paper proposes the Adaptive EfficientNet, a novel framework that integrates dynamic compound scaling to achieve resource-optimal deep learning. We present a methodology for adjusting the EfficientNet's scaling parameters (width, depth, and resolution coefficients) on-the-fly or at deployment time based on real-time resource availability or predefined budget constraints, such as target latency, computational budget (FLOPS), or memory footprint. Our approach leverages a resource-aware optimization strategy to find the most efficient network configuration for a given set of constraints, thereby maximizing performance under limited resources. We demonstrate through extensive experimental analysis that Adaptive EfficientNet significantly outperforms statically scaled models by achieving comparable or superior accuracy with substantially reduced computational requirements or by providing improved accuracy within strict resource envelopes. This work paves the way for more flexible and efficient deployment of deep learning models in heterogeneous and resource-constrained environments, offering a crucial advancement towards ubiquitous and sustainable AI.
Related Topics To Compare & Contrast
- Type
- article
- Landing Page
- https://doi.org/10.5281/zenodo.17822377
- OA Status
- green
- OpenAlex ID
- https://openalex.org/W7108645210