ACO-Pruning for Deep Neural Networks: A Case Study in CNNs
Renato Sellaro Dorighello, Myriam Delgado, Ricardo Lüders, Daniel Fernando Pigatto · Proceedings of the Genetic and Evolutionary Computation Conference Companion · 2024
Deep Neural Networks (DNNs) are successful in several tasks, mainly due to their ability to process a large volume of data, given their huge number of parameters and computational operations. Larger and deeper models have been developed to improve their performance with an increasing computational cost. Pruning algorithms are strategies necessary to mitigate the computational burden and achieve better performance by eliminating parts of the network structure while maintaining good training and testing results. Dynamic network pruning increases performance through online choices of inference paths depending on various inputs. This work proposes a new Ant Colony Optimization Pruning (ACO-P) algorithm for dynamic pruning based on swarm intelligence to compress the model without jeopardizing accuracy. We validate ACO-P with a CNN model on the MNIST dataset by comparing it with a baseline pruner that uses random choices, and a well-established dynamic pruning method based on a secondary neural network. The results show that our proposal is a computationally more efficient alternative, capable of achieving higher pruning rates.