Towards Adaptive Deep Learning: Model Elasticity via Prune-and-Grow CNN Architectures

TL;DR

This work tackles the challenge of deploying CNNs on resource-constrained devices by introducing a prune-and-grow approach that embeds nested subnetworks within a single model. Leveraging dependency-aware pruning via DepGraph, the method creates a core pruned model and a rebuilding pipeline that reinserts pruned components to regain full capacity without retraining from scratch. Across VGG-16, AlexNet, ResNet-20, and ResNet-56 on CIFAR-10, the approach demonstrates significant reductions in parameters and model size, while maintaining or even surpassing baseline accuracy after rebuilding, especially with iterative pruning. The resulting elastic CNNs show promise for robust, on-device deployment across heterogeneous hardware, enabling runtime adaptation without extensive retraining.

Abstract

Deploying deep convolutional neural networks (CNNs) on resource-constrained devices presents significant challenges due to their high computational demands and rigid, static architectures. To overcome these limitations, this thesis explores methods for enabling CNNs to dynamically adjust their computational complexity based on available hardware resources. We introduce adaptive CNN architectures capable of scaling their capacity at runtime, thus efficiently balancing performance and resource utilization. To achieve this adaptability, we propose a structured pruning and dynamic re-construction approach that creates nested subnetworks within a single CNN model. This approach allows the network to dynamically switch between compact and full-sized configurations without retraining, making it suitable for deployment across varying hardware platforms. Experiments conducted across multiple CNN architectures including VGG-16, AlexNet, ResNet-20, and ResNet-56 on CIFAR-10 and Imagenette datasets demonstrate that adaptive models effectively maintain or even enhance performance under varying computational constraints. Our results highlight that embedding adaptability directly into CNN architectures significantly improves their robustness and flexibility, paving the way for efficient real-world deployment in diverse computational environments.

Figures (11)

• Figure 1: Overview of the iterative pruning and rebuilding process. The model is progressively compressed via structured pruning steps and later reconstructed in reverse using preserved weight metadata. pruningfiltersefficientconvnets
• Figure 2: Dependency-aware pruning ensures consistency across layers with structural dependencies. For example, pruning Conv2 also requires pruning Conv1, BN1, and BN2 within the block. The Dependency Graph tracks such relationships automatically. Source: depGraph
• Figure 3: Illustration of model freezing and rebuilding
• Figure 4: Overview of the pruning and rebuilding process for adaptivity. Structured pruning removes entire filters (shown as dotted circles), resulting in a smaller core model. Later, previously pruned filters are reinserted into the architecture during the rebuilding phase, allowing the model to adapt to increased resource availability.
• Figure 5: Overview of the iterative pruning and rebuilding process. The model is progressively compressed via structured pruning steps and later reconstructed in reverse using preserved weight metadata.