Abstract
In the dynamic landscape of Artificial Intelligence (AI) advancements, particularly in the development of compact and highly efficient models for space-constrained environments, the strategic sparsification of neural networks takes center stage. In this work, we investigate creating, training, and evaluating Convolutional Neural Network (CNN), DenseNet, and ResNet models taking advantage of sparse neural networks with the help of the Sparse Evolutionary Training (SET) approach. This paper extends the existing framework, while also considering critical performance metrics, such as accuracy, precision, recall, inference time, execution time, memory usage, and energy consumption. The proposed script facilitates exploration across diverse model architectures, sparsity levels, quantization options, and the number of training epochs, alongside a recording of these performance metrics throughout both the training and inference phases. We analyse and evaluate the performance of the chosen models and parameters on a classic Malaria dataset. Results confirm that highly sparse ANNs (e.g., 70%) can achieve results comparable to their fully-connected counterparts while allowing for smaller models and possibly more energy-efficient systems. Our contribution advances the ongoing discussion on optimizing embedded systems for efficient AI applications.