<ul><li>A new thesis explores methods for enabling CNNs to dynamically adjust their computational complexity based on available hardware resources.</li><li>The thesis introduces adaptive CNN architectures that can scale their capacity at runtime, balancing performance and resource utilization efficiently.</li><li>A structured pruning and dynamic re-construction approach is proposed to create nested subnetworks within a single CNN model, allowing for dynamic configuration switching without retraining.</li><li>Experiments on various CNN architectures like VGG-16, AlexNet, ResNet-20, and ResNet-56 show that adaptive models maintain or enhance performance under varying computational constraints, improving robustness and flexibility for deployment in diverse environments.</li></ul>

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

Discover more