We introduce a novel neural network architecture for encoding and synthesis of 3D shapes, particularly their structures. Our key insight is that 3D shapes are effectively characterized by their hierarchical organizati...
详细信息
We introduce a novel neural network architecture for encoding and synthesis of 3D shapes, particularly their structures. Our key insight is that 3D shapes are effectively characterized by their hierarchical organization of parts, which reflects fundamental intra-shape relationships such as adjacency and symmetry. We develop a recursive neural net (RvNN) based autoencoder to map a flat, unlabeled, arbitrary part layout to a compact code. The code effectively captures hierarchical structures of man-made 3D objects of varying structural complexities despite being fixed-dimensional: an associated decoder maps a code back to a full hierarchy. The learned bidirectional mapping is further tuned using an adversarial setup to yield a generative model of plausible structures, from which novel structures can be sampled. Finally, our structure synthesis framework is augmented by a second trained module that produces fine-grained part geometry, conditioned on global and local structural context, leading to a full generative pipeline for 3D shapes. We demonstrate that without supervision, our network learns meaningful structural hierarchies adhering to perceptual grouping principles, produces compact codes which enable applications such as shape classification and partial matching, and supports shape synthesis and interpolation with significant variations in topology and geometry.
With computers and the Internet being essential in everyday life, malware poses serious and evolving threats to their security, making the detection of malware of utmost concern. Accordingly, there have been many rese...
详细信息
With computers and the Internet being essential in everyday life, malware poses serious and evolving threats to their security, making the detection of malware of utmost concern. Accordingly, there have been many researches on intelligent malware detection by applying data mining and machine learning techniques. Though great results have been achieved with these methods, most of them are built on shallow learning architectures. Due to its superior ability in feature learning through multilayer deep architecture, deep learning is starting to be leveraged in industrial and academic research for different applications. In this paper, based on the Windows application programming interface calls extracted from the portable executable files, we study how a deep learning architecture can be designed for intelligent malware detection. We propose a heterogeneous deep learning framework composed of an autoencoder stacked up with multilayer restricted Boltzmann machines and a layer of associative memory to detect newly unknown malware. The proposed deep learning model performs as a greedy layer-wise training operation for unsupervised feature learning, followed by supervised parameter fine-tuning. Different from the existing works which only made use of the files with class labels (either malicious or benign) during the training phase, we utilize both labeled and unlabeled file samples to pre-train multiple layers in the heterogeneous deep learning framework from bottom to up for feature learning. A comprehensive experimental study on a real and large file collection from Comodo Cloud Security Center is performed to compare various malware detection approaches. Promising experimental results demonstrate that our proposed deep learning framework can further improve the overall performance in malware detection compared with traditional shallow learning methods, deep learning methods with homogeneous framework, and other existing anti-malware scanners. The proposed heterogeneous
In this paper, we analyze the role of hidden bias in representational efficiency of the Gaussian-Bipolar Restricted Boltzmann Machines (GBPRBMs), which are similar to the widely used Gaussian-Bernoulli RBMs. Our exper...
详细信息
In this paper, we analyze the role of hidden bias in representational efficiency of the Gaussian-Bipolar Restricted Boltzmann Machines (GBPRBMs), which are similar to the widely used Gaussian-Bernoulli RBMs. Our experiments show that hidden bias plays an important role in shaping of the probability density function of the visible units. We define hidden entropy and propose it as a measure of representational efficiency of the model. By using this measure, we investigate the effect of hidden bias on the hidden entropy and provide a full analysis of the hidden entropy as function of the hidden bias for small models with up to three hidden units. We also provide an insight into understanding of the representational efficiency of the larger scale models. Furthermore, we introduce Normalized Empirical Hidden Entropy (NEHE) as an alternative to hidden entropy that can be computed for large models. Experiments on the MNIST, CIFAR-10 and Faces data sets show that NEHE can serve as measure of representational efficiency and gives an insight on minimum number of hidden units required to represent the data. (c) 2018 Elsevier Ltd. All rights reserved.
Through-wall human detection has vital and widely used applications for anti-terrorism, anti-explosion, and post-disaster relief. The through-wall human-target recognition using ultra-wideband radar-based technology w...
详细信息
Through-wall human detection has vital and widely used applications for anti-terrorism, anti-explosion, and post-disaster relief. The through-wall human-target recognition using ultra-wideband radar-based technology was established in recent research. With the recent development of deep learning algorithms, classification algorithms have demonstrated a dynamic aptitude to learn important characteristics of the dataset by utilizing only a few sample sets. This paper focuses on studying the detection of a human target's status behind wall in small sample conditions. In the deep learning network model, the autoencoder algorithm is chosen here to classify and identify human targets behind walls. Through automatic acquiring of the knowledge of inherent characteristics in the data, the autoencoder algorithm can extract the concise data-feature representations. Based on the autoencoder network, we add the denoising encoder and sparsity constraints to extract more efficient feature representations, thereby improving the classification and identification rates. In this paper, we classify and identify the behind-wall human-target states separately under single and multiple sensors under a small-sample condition, and then compare the results with those of other classification algorithms. The results illustrate that the use of multiple sensors is more effective than the use of a single sensor and that the adopted autoencoder algorithm enables more effective detection of human targets behind walls than other algorithms.
In this paper, we propose to use the denoising autoencoder to generate robust feature representations for emotion recognition. In our method, the input of the denoising autoencoder is the normalized static feature set...
详细信息
ISBN:
(纸本)9781629934433
In this paper, we propose to use the denoising autoencoder to generate robust feature representations for emotion recognition. In our method, the input of the denoising autoencoder is the normalized static feature set (state-of-the-art features for emotion recognition). This input is mapped to two hidden representations: one is to capture the neutral information from the input, and the other one is used to extract emotional information. Model parameters are learned by minimizing the squared error between the original and the reconstructed input. After pre-training and fine-tuning, we use the hidden representation as features in the SVM model for emotion classification. Our experimental results show significant performance improvement compared to using the static features.
With an increasing amount of data in intelligent transportation s)stems, methods are needed to automatically extract general representations that accurately predict not only known tasks but also similar tasks that can...
详细信息
With an increasing amount of data in intelligent transportation s)stems, methods are needed to automatically extract general representations that accurately predict not only known tasks but also similar tasks that can emerge in the future. Creation of low-dimensional representations can be unsupervised or can exploit various labels in multi-task learning (when goal tasks are known) or transfer learning (when they are not) settings. Finding a general, low-dimensional representation suitable for multiple tasks is an important step toward knowledge discovery in aware intelligent transportation systems. This paper evaluates several approaches mapping high-dimensional sensor data from Volvo trucks into a low-dimensional representation that is useful for prediction. Original data are bivariate histograms, with two types-turbocharger and engine-considered. Low-dimensional representations were evaluated in a supervised fashion by mean equal error rate (EER) using a random forest classifier on a set of 27 1-vs-Rest detection tasks. Results from unsupervised learning experiments indicate that using an autoencoder to create an intermediate representation, followed by t-distributed stochastic neighbor embedding, is the most effective way to create low-dimensional representation of the original bivariate histogram. Individually, t-distributed stochastic neighbor embedding offered best results for 2-D or 3-D and classical autoencoder for 6-D or 10-D representations. Using multi-task learning, combining unsupervised and supervised objectives on all 27 available tasks, resulted in 10-D representations with a significantly lower EER compared to the original 400-D data. In transfer learning setting, with topmost diverse tasks used for representation learning, 10-D representations achieved EER comparable to the original representation.
The ability to generate novel, diverse, and realistic 3D shapes along with associated part semantics and structure is central to many applications requiring high-quality 3D assets or large volumes of realistic trainin...
详细信息
The ability to generate novel, diverse, and realistic 3D shapes along with associated part semantics and structure is central to many applications requiring high-quality 3D assets or large volumes of realistic training data. A key challenge towards this goal is how to accommodate diverse shape variations, including both continuous deformations of parts as well as structural or discrete alterations which add to, remove from, or modify the shape constituents and compositional structure. Such object structure can typically be organized into a hierarchy of constituent object parts and relationships, represented as a hierarchy of n-ary graphs. We introduce STRUCTURENET, a hierarchical graph network which (i) can directly encode shapes represented as such n-ary graphs, (ii) can be robustly trained on large and complex shape families, and (iii) be used to generate a great diversity of realistic structured shape geometries. Technically, we accomplish this by drawing inspiration from recent advances in graph neural networks to propose an order-invariant encoding of n-ary graphs, considering jointly both part geometry and inter-part relations during network training. We extensively evaluate the quality of the learned latent spaces for various shape families and show significant advantages over baseline and competing methods. The learned latent spaces enable several structure-aware geometry processing applications, including shape generation and interpolation, shape editing, or shape structure discovery directly from un-annotated images, point clouds, or partial scans.
Effective feature representation is crucial to view-based 3D object retrieval (V3OR). Most previous works employed hand-crafted features to represent the views of each object. Although deep learning based methods has ...
详细信息
Effective feature representation is crucial to view-based 3D object retrieval (V3OR). Most previous works employed hand-crafted features to represent the views of each object. Although deep learning based methods has shown its excellent performance in many vision tasks, it is hard to get excellent performance for unsupervised 3D object retrieval. In this paper, we propose to combine the off-the-shelf deep model and graph model to retrieve unseen objects. By employing the powerful deep classification models which are trained from millions of images, we obtain significant improvements compared with state of the art methods. We validate the effectiveness of the ready CNN from other domains that can greatly facilitate the representative ability of objects' views. In addition, we analyze the representative abilities of different fully connected layers for V3OR, and propose to employ multigraph learning to fuse the deep features of different layers. The autoencoder is then explored to improve the retrieval speed to a large extent. Experiments on two popular datasets are carried out to demonstrate the effectiveness of the proposed method.
At present, the semantic information segmentation algorithms mainly include FCN (Fully Convolutional Network), PSPNet (Pyramid Scene Parsing Network), Deeplab and so on. In view of the inadequate results of features e...
详细信息
At present, the semantic information segmentation algorithms mainly include FCN (Fully Convolutional Network), PSPNet (Pyramid Scene Parsing Network), Deeplab and so on. In view of the inadequate results of features extracted by these algorithms from RGB image, a hybrid fully convolutional autoencoder neural network (HFCAN) structure, which introduces fully convolutional neural network and stacked sparse autoencoder, is proposed in this paper. Using the FCN to generate the thermal high-dimensional feature map of the shelf commodity, and then performing the up-sampling operation on the segmented feature map. During the up-sampling operation, the convolution features are refined by the stacked sparse autoencoder (SAE), and the image boundary details are retained, so that the classification results are more accurate. The experimental results show that the hybrid fully convolutional autoencoder model proposed in this paper can not only shorten the training time and testing time of neural network by nearly 50, but also improve the accuracy of shelf commodity identification by nearly 95.
Quantum many-body systems realize many different phases of matter characterized by their exotic emergent phenomena. While some simple versions of these properties can occur in systems of free fermions, their occurrenc...
详细信息
Quantum many-body systems realize many different phases of matter characterized by their exotic emergent phenomena. While some simple versions of these properties can occur in systems of free fermions, their occurrence generally implies that the physics is dictated by an interacting Hamiltonian. The interaction distance has been successfully used to quantify the effect of interactions in a variety of states of matter via the entanglement spectrum [C. J. Turner, K. Meichanetzidis, Z. Papic and J. K. Pachos, Nat. Commun. 8 (2017) 14926, Phys. Rev. B97 (2018) 125104]. The computation of the interaction distance reduces to a global optimization problem whose goal is to search for the free-fermion entanglement spectrum closest to the given entanglement spectrum. In this work, we employ techniques from machine learning in order to perform this same task. In a supervised learning setting, we use labeled data obtained by computing the interaction distance and predict its value via linear regression. Moving to a semi-supervised setting, we train an autoencoder to estimate an alternative measure to the interaction distance, and we show that it behaves in a similar manner.
暂无评论