Data embedding is used in many machine learning applications to create low-dimensional feature representations, which preserves the structure of data points their original space. In this paper, we examine scenario a heterogeneous network with nodes and content various types. Such networks are notoriously difficult mine because bewildering combination contents structures. The creation multidimensional such opens door use wide variety off-the-shelf mining techniques for data. Despite...

Automatically annotating concepts for video is a key to semantic-level browsing, search and navigation. The research on this topic evolved through two paradigms. first paradigm used binary classification detect each individual concept in set. It achieved only limited success, as it did not model the inherent correlation between concepts, e.g., urban building. second added step top of detectors fuse multiple concepts. However, its performance varies because errors incurred detection can...

The long short-term memory (LSTM) neural network is capable of processing complex sequential information since it utilizes special gating schemes for learning representations from input sequences. It has the potential to model any time-series or data, where current hidden state be considered in context past states. This property makes LSTM an ideal choice learn dynamics various actions. Unfortunately, conventional LSTMs do not consider impact spatio-temporal corresponding given salient...

Learning-based video annotation is a promising approach to facilitating retrieval and it can avoid the intensive labor costs of pure manual annotation. But frequently encounters several difficulties, such as insufficiency training data curse dimensionality. In this paper, we propose method named optimized multigraph-based semi-supervised learning (OMG-SSL), which aims simultaneously tackle these difficulties in unified scheme. We show that various crucial factors annotation, including...

Meta-learning approaches have been proposed to tackle the few-shot learning problem. Typically, a meta-learner is trained on variety of tasks in hopes being generalizable new tasks. However, generalizability could be fragile when it over-trained existing during meta-training phase. In other words, initial model too biased towards adapt tasks, especially only very few examples are available update model. To avoid and improve its generalizability, we propose novel paradigm Task-Agnostic...

Differentiable architecture search (DARTS) provided a fast solution in finding effective network architectures, but suffered from large memory and computing overheads jointly training super-network searching for an optimal architecture. In this paper, we present novel approach, namely, Partially-Connected DARTS, by sampling small part of to reduce the redundancy exploring space, thereby performing more efficient without comprising performance. particular, perform operation subset channels...

Stock prices are formed based on short and/or long-term commercial and trading activities that reflect different frequencies of patterns. However, these patterns often elusive as they affected by many uncertain political-economic factors in the real world, such corporate performances, government policies, even breaking news circulated across markets. Moreover, time series stock non-stationary non-linear, making prediction future price trends much challenging. To address them, we propose a...

In this paper, we present a simple and modularized neural network architecture, named interleaved group convolutional networks (IGCNets). The main point lies in novel building block, pair of two successive convolutions: primary convolution secondary convolution. convolutions are complementary: (i) the on each partition is spatial convolution, while point-wise convolution; (ii) channels same come from different partitions. We discuss one representative advantage: Wider than regular with...

In this article, we exploit the problem of annotating a large-scale image corpus by label propagation over noisily tagged web images. To annotate images more accurately, propose novel k NN-sparse graph-based semi-supervised learning approach for harnessing labeled and unlabeled data simultaneously. The sparse graph constructed datum-wise one-vs- NN reconstructions all samples can remove most semantically unrelated links among data, thus it is robust discriminative than conventional graphs....

Representation learning with small labeled data have emerged in many problems, since the success of deep neural networks often relies on availability a huge amount that is expensive to collect. To address it, efforts been made training sophisticated models few an unsupervised and semi-supervised fashion. In this paper, we will review recent progresses these two major categories methods. A wide spectrum be categorized big picture, where show how they interplay each other motivate explorations...

Traffic forecasting has emerged as a core component of intelligent transportation systems. However, timely accurate traffic forecasting, especially long-term still remains an open challenge due to the highly nonlinear and dynamic spatial-temporal dependencies flows. In this paper, we propose novel paradigm Spatial-Temporal Transformer Networks (STTNs) that leverages dynamical directed spatial long-range temporal improve accuracy forecasting. Specifically, present new variant graph neural...

Human can well recognize images of novel categories just after browsing few examples these categories. One possible reason is that they have some external discriminative visual information about from their prior knowledge. Inspired this, we propose a Knowledge Transfer Network architecture (KTN) for few-shot image recognition. The proposed KTN model jointly incorporates feature learning, knowledge inferring and classifier learning into one unified framework optimal compatibility. First, the...

In recent years, deep networks have been successfully applied to model image concepts and achieved competitive performance on many data sets. spite of impressive performance, the conventional can be subjected decayed if we insufficient training examples. This problem becomes extremely severe for with powerful representation structure, making them prone over fitting by capturing nonessential or noisy information in a small set. this paper, address challenge, will develop novel network capable...

Human motion prediction aims to generate future motions based on the observed human motions. Witnessing success of Recurrent Neural Networks (RNN) in modeling sequential data, recent works utilize RNNs model human-skeleton sequence and predict However, these methods disregard existence spatial coherence among joints temporal evolution skeletons, which reflects crucial characteristics spatiotemporal space. To this end, we propose a novel Skeleton-Joint Co-Attention (SC-RNN) capture joints,...

The success of deep neural networks often relies on a large amount labeled examples, which can be difficult to obtain in many real scenarios. To address this challenge, unsupervised methods are strongly preferred for training without using any data. In paper, we present novel paradigm representation learning by Auto-Encoding Transformation (AET) contrast the conventional Data (AED) approach. Given randomly sampled transformation, AET seeks predict it merely from encoded features as...

In this work, we aim to address the problem of human interaction recognition in videos by exploring long-term inter-related dynamics among multiple persons. Recently, Long Short-Term Memory (LSTM) has become a popular choice model individual dynamic for single-person action due its ability capture temporal motion information range. However, most existing LSTM-based methods focus only on capturing simply combining all individuals or modeling them as whole. Such neglect how interactions change...

Representation learning has significantly been developed with the advance of contrastive methods. Most those methods are benefited from various data augmentations that carefully designated to maintain their identities so images transformed same instance can still be retrieved. However, designed transformations limited us further explore novel patterns exposed by other transformations. Meanwhile, as shown in our experiments, direct for stronger augmented not learn representations effectively....

Unsupervised domain adaptation (UDA) aims to transfer knowledge learned from a fully-labeled source different unlabeled target domain. Most existing UDA methods learn domain-invariant feature representations by minimizing distances across domains. In this work, we build upon contrastive self-supervised learning align features so as reduce the discrepancy between training and testing sets. Exploring same set of categories shared both domains, introduce simple yet effective framework CDCL, for...

Recently, deep convolution neural networks (CNNs) steered face super-resolution methods have achieved great progress in restoring degraded facial details by joint training with priors. However, these some obvious limitations. On the one hand, multi-task learning requires additional marking on dataset, and introduced prior network will significantly increase computational cost of model. other limited receptive field CNN reduce fidelity naturalness reconstructed images, resulting suboptimal...

Deep learning-based models have been shown to outperform human beings in many computer vision tasks with massive available labeled training data learning. However, humans an amazing ability easily recognize images of novel categories by browsing only a few examples these categories. In this case, few-shot learning comes into being make machines learn from extremely limited examples. One possible reason why can well concepts quickly and efficiently is that they sufficient visual semantic...

In real world, an image is usually associated with multiple labels which are characterized by different regions in the image. Thus classification naturally posed as both a multi-label learning and multi-instance problem. Different from existing research has considered these two problems separately, we propose integrated (MLMIL) approach based on hidden conditional random fields (HCRFs), simultaneously captures connections between semantic regions, correlations among single formulation. We...

In this paper, we exploit the problem of inferring images' semantic concepts from community-contributed images and their associated noisy tags. To infer more accurately, propose a novel sparse graph-based semi-supervised learning approach for harnessing labeled unlabeled data simultaneously. The graph constructed by datum-wise one-vs-all reconstructions all samples can remove most concept-unrelated links among data, thus is robust discriminative than conventional graphs. More importantly, an...

Cu2ZnSnS4 (CZTS) and its related materials such as Cu2ZnSnSe4 (CZTSe) Cu2ZnSn(S,Se)4 (CZTSSe) have attracted considerable attention an absorber material for thin film solar cells due to the non-toxicity, elemental abundance, large production capacity of their constituents. Despite similarities between CZTS-based Cu(In,Ga)Se2(CIGS), record efficiency remains significantly lower than that CIGS cells. Considering difference two lies in choice material, cause can be isolated issues associated...