Core Concepts
Recombining neural networks through stitching can lead to novel trade-offs between performance and computational cost.
Abstract
The content delves into the concept of stitching for neuroevolution, focusing on recombining deep neural networks without compromising their integrity. The authors propose a method that allows for the creation of new networks by leveraging pre-trained models and introducing stitching layers. This approach aims to find a balance between network performance and computational efficiency. The article discusses the challenges of recombining neural networks, the process of efficient model stitching, and the study of offspring space. Experimental setups and results from tasks like ImageNet classification and Semantic Segmentation on VOC dataset are detailed, showcasing the effectiveness of the proposed approach in finding networks that outperform or dominate parent networks.
Structure:
Introduction to Neuroevolution
Traditional vs. modern approaches
Efficient Model Stitching
Overcoming compatibility issues in recombination
Studying Offspring Space
Investigating potential trade-offs in offspring networks
Experimental Setup and Results
Tasks, population sizes, evaluation budgets, and outcomes
Comparison of Approaches
GA vs GOMEA vs LK-GOMEA
Found Networks Analysis
Fronts obtained on validation vs test sets
Calibration Analysis
Expected Calibration Error (ECE) for best networks found
Conclusion and Acknowledgments
Stats
"The resulting network has 154 matches."
"This transformation introduces stitching layers to transform the output of layer B into what the output of layer A is expected to be."
"The resulting network has 206 matches."
Quotes
"Our approach enables finding networks that represent novel trade-offs between performance and computational cost."
"Creating this supernetwork took a total of 1415.52 seconds."