Deep Learning Applications

Abstract

The manuscript presents three of my deep learning projects: mDL-ArtTranfer – Deep Learning Art Transfer using Multiple AIs (https://github.com/muntisa/mDL-ArtTransfer), CNN4Polyps - Colonoscopy polyp detection with Convolutional Neural Networks (https://github.com/muntisa/Colonoscopy-polyps-detection-with-CNNs), Deep-Politics - Prediction of Spanish Political Affinity with Deep Neural Nets: Socialist vs People's Party (PSOE vs PP, https://github.com/muntisa/Deep-Politics). mDL-ArtTranfer is a mix of adapted scripts using three AI algorithms from fchollet, anishathalye, and ShafeenTejani (GitHub users). Thus, using content images and style pictures, three versions of art transfer will be apply with only one single call. CNN4Polyps represents the first open GitHub repository for polyp detection and localization into colonoscopy images. The use of small CNNs with 2-3 convolutions, in only 2 minutes with GPU Nvidia Titan Xp, will generate a model with of 92%. The VGG16 transfer learning is no improving this accuracy. The fine tuning of the last convolutional block and the full connected layer of the pre-trained Imagenet VGG16 will generate an accuracy over 98%. Deep-Politics is using the politician’s portrait to predict the affinity for two political parties in Spain. Both small CNNs with augmented data and VGG16 transfer learning without data augmentation can generate models with accuracy over 80%. The VGG16 fine tuning of the last two convolutional blocks and the full connected layer will raise the accuracy to 85%.