Plataforma automática para compressão consciente de modelos de aprendizagem profunda baseada em poda seguida de quantização

Abstract

The rapid advancement of deep learning applications has significantly impacted various sectors. However, the size and complexity of neural models pose challenges for deployment in systems with limited computational resources. This work presents an automated platform that integrates pruning and quantization techniques during the training process of deep neural networks to optimize both accuracy and computational efficiency. Experiments conducted on the CIFAR-10 dataset demonstrated the effectiveness of the proposed approach, achieving up to 80% reduction in memory size while maintaining accuracy levels comparable to the state-of-the-art, with minimal losses of 2-3%. Analysis of weight distribution before and after compression revealed adaptation patterns of the network during compression-based training, providing valuable insights into model behavior under different compression regimes.