International Journal of Machine Intelligence and Sensory Signal Processing (5 papers in press)
Designing Ensemble Learning Algorithms using Kernel Methods
by Fayao Liu
Abstract: Ensemble methods such as boosting combine multiple learners to obtain better prediction than could be obtained from any individual learner. Here we propose a principled framework for directly constructing ensemble learning methods from kernel methods. Unlike previous studies showing the equivalence between boosting and support vector machines (SVMs), which need a translation procedure, we show that it is possible to design boosting-like procedure to solve the SVM optimization problems. In other words, it is possible to design ensemble methods directly from SVM without any middle procedure. This finding not only enables us to design new ensemble learning methods directly from kernel methods, but also makes it possible to take advantage of those highly-optimized fast linear SVM solvers for ensemble learning. The resulted model is as effective as kernel methods while being as efficient as ensemble methods. We exemplify this framework for designing new binary and multi-class classification ensemble learning as well as a new quantile regression ensemble learning method. Experimental results demonstrate the flexibility and usefulness of the proposed framework.
Keywords: Kernel; Support vector machines; Ensemble learning; Column generation; Multi-class classification; Quantile regression.
Deep Learning in Digital Marketing: Brand Detection and Emotion Recognition
by Bernardete Ribeiro, Gonçalo Oliveira, Ana Laranjeira, Joel Arrais
Abstract: In this paper we explore deep learning in two case studies, brand detection and emotion recognition, that can have a prominent role in the marketing industry. The information that can be retrieved from brand detection technology and emotion recognition can increase competitiveness via the activity on digital social interactions between branding and the end-consumer. rnDeep learning is extremely good for visual feature extraction from images, audio signals or text which makes it very attractive to be used today. Given that specific datasets of logo images and emotion facial expression images can be easily processed by deep neural networks without the need to extract manually hand-crafted features, much higher performance than traditional methods can be obtained from these models. In particular, Convolutional Neural Networks (CNN) are becoming the current state of art in many image processing problems.rnIn the first case study, we build a graphic logo detection by using a Fast Region-based Convolutional Network (FRCN). This method looks for region proposals in the logo image, since the logo is often present in small sizes and partially occluded. This avoids a full search in the image and improves the object detection. Furthermore, instead of building the CNNs from the scratch which would be mostly prohibitive, transfer learning and data augmentation were used and have shown to excel previous approaches.rnIn the second case study, we present a robust way of facial emotions recognition by introducing an improved version of the classic CNN - LeNet-5. Despite the net simplicity, it was found to be better suited for the system constraints (dataset dimension, faces size and composition) by comparing its performance with deeper networks such as GoogleNet and AlexNet. Although a video tracker could successfully trace some of the facial expressions further improvements are still needed.Deep Learning; Brand Detection; Emotion Recognition
Keywords: Deep Learning; Convolutional Neural Networks; Brand Detection; Emotion Recognition.
Semi-Supervised Feature Selection with Sparse Representation for Hyperspectral Image Classification
by Yanyan Zhang, Shiguo Chen, Cailing Wang, Zhisong Pan, Daoqiang Zhang
Abstract: Dimensionality reduction is one of the most important steps for remotely sensed hyperspectral image classification. Feature selection as a kind of dimensionality reduction has attracted great attentions in the recent decades. In this paper, we proposed a novel feature selection method for hyperspectral image classification based on semi-supervised learning and sparsity score (or briefly called Semi-supervised Sparsity Score (Semi-SS)). In Semi-SS, the pairwise constraints instead of class labels are used as the supervision information. Furthermore, the features chosen by Semi-SS have the ability to reconstruct the original data and sparsity preserving. Experiments conducted on two famous hyperspectral data sets illustrate that the proposed algorithm is remarkably effective in comparison to the existing feature selection methods.
Keywords: Hyperspectral Image Classification; Semi-Supervised Feature Selection; Sparse Representation; Pairwise Constraints.
A Weighted Hybrid Training Algorithm of Neural Networks for Robust Data Regression
by Feilong Cao, Sifang Che, Jianwei Zhao
Abstract: Hybrid full memory Broyden-Fletcher-Golfarb-Shanno (BFGS) algorithm (HFM)rnis an effective technique for training feed-forward neural networks. It is superiority in training speed compared to classic second-order gradient methods. However, for training data with outliers, the performance of HFM is extremely affected. This paper addresses the robustness of HFM for regression with outliers, and proposes a new algorithm, named weighted regularization hybrid training algorithm (WRHFM), to solve this drawback. The main idea for this is to weight the error variable of regularization hybrid training algorithm (RHFM) by weighting factors, and to establish a robust model against outliers. The experiments are conducted on both function approximation and real-world data, which show that the proposed approach has the well robustness performance compared with classical back propagation algorithm (BP), supported vector machine (SVM), HFM, and RHFM in handling data with outliers.
Keywords: Feed-forward neural networks; hybrid full memory; alternative
optimization; robust data regression.
On Developing Sugeno Fuzzy Measure Densities in Problems of Face Recognition
by Paweł Karczmarek, Adam Kiersztyn, Witold Pedrycz
Abstract: Fuzzy measures and Choquet integral are efficient aggregation operators utilized intensively in decision-making theory. To obtain sound classification results based on a family of classifiers, the parameters of the fuzzy measure (especially socalled fuzzy densities) have to be determined. In this study, we propose a method based on Particle Swarm Optimization (PSO) and discuss in detail a new concept of a so-called positive and negative optimization to fully utilize specific properties of classifiers to carry out efficient classification. A suite of experiments is conducted to illustrate this approach and discuss its scope of applicability.
Keywords: fuzzy measure; Choquet integral; Particle Swarm Optimization; face recognition.