Person: Pajares Martínsanz, Gonzalo
Universidad Complutense de Madrid
Faculty / Institute
Ingeniería del Software e Inteligencia Artificial
Lenguajes y Sistemas Informáticos
Now showing 1 - 2 of 2
PublicationFilter pruning for convolutional neural networks in semantic image segmentation(Elsevier, 2024-01) López González, Clara Isabel; Gascó, Esther; Barrientos Espillco, Fredy; Besada Portas, Eva; Pajares Martínsanz, GonzaloThe remarkable performance of Convolutional Neural Networks (CNNs) has increased their use in real-time systems and devices with limited resources. Hence, compacting these networks while preserving accuracy has become necessary, leading to multiple compression methods. However, the majority require intensive iterative procedures and do not delve into the influence of the used data. To overcome these issues, this paper presents several contributions, framed in the context of explainable Artificial Intelligence (xAI): (a) two filter pruning methods for CNNs, which remove the less significant convolutional kernels; (b) a fine-tuning strategy to recover generalization; (c) a layer pruning approach for U-Net; and (d) an explanation of the relationship between performance and the used data. Filter and feature maps information are used in the pruning process: Principal Component Analysis (PCA) is combined with a next-convolution influence-metric, while the latter and the mean standard deviation are used in an importance score distribution-based method. The developed strategies are generic, and therefore applicable to different models. Experiments demonstrating their effectiveness are conducted over distinct CNNs and datasets, focusing mainly on semantic segmentation (using U-Net, DeepLabv3+, SegNet, and VGG-16 as highly representative models). Pruned U-Net on agricultural benchmarks achieves 98.7% parameters and 97.5% FLOPs drop, with a 0.35% gain in accuracy. DeepLabv3+ and SegNet on CamVid reach 46.5% and 72.4% parameters reduction and a 51.9% and 83.6% FLOPs drop respectively, with almost no decrease in accuracy. VGG-16 on CIFAR-10 obtains up to 86.5% parameter and 82.2% FLOPs decrease with a 0.78% accuracy gain. PublicationLayer factor analysis in convolutional neural networks for explainability(Elsevier, 2024-01) López González, Clara Isabel; Gómez Silva, María José; Besada Portas, Eva; Pajares Martínsanz, GonzaloExplanatory methods that focus on the analysis of the features encoded by Convolutional Neural Networks (CNNs) are of great interest, since they help to understand the underlying process hidden behind the black-box nature of these models. However, to explain the knowledge gathered in a given layer, they must decide which of the numerous filters to study, further assuming that each of them corresponds to a single feature. This, coupled with the redundancy of information, makes it difficult to ensure that the relevant characteristics are being analyzed. The above represents an important challenge and defines the aim and scope of our proposal. In this paper we present a novel method, named Explainable Layer Factor Analysis for CNNs (ELFA-CNNs), which models and describes with quality convolutional layers relying on factor analysis. Regarding contributions, ELFA obtains the essential underlying features, together with their correlation with the original filters, providing an accurate and well-founded summary. Through the factorial parameters we gain insights about the information learned, the connections between channels, and the redundancy of the layer, among others. To provide visual explanations in a similarly way to other methods, two additional proposals are made: a) Essential Feature Attribution Maps (EFAM) and b) intrinsic features inversion. The results prove the effectiveness of the developed general methods. They are evaluated in different CNNs (VGG-16, ResNet-50, and DeepLabv3+) on generic datasets (CIFAR-10, imagenette, and CamVid). We demonstrate that convolutional layers adequately fit a factorial model thanks to the new metrics presented for factor and fitting residuals (D1, D>, and Res, derive from covariance matrices). Moreover, knowledge about the deep image representations and the learning process is acquired, as well as reliable heat maps highlighting regions where essential features are located. This study effectively provides an explainable approach that can be applied to different CNNs and over different datasets.