This paper proposes a novel sparse variant of auto-encoders as building block to pre-train deep neural networks. Compared with through KL-divergence, our method requires fewer hyper-parameters and the sparsity level hidden units can be learnt automatically. We have compared several other unsupervised leaning algorithms on benchmark databases. The satisfactory classification accuracy (97.92% MNIST 87.29% NORB) achieved by 2-hidden-layer network pre-trained using algorithm, whole training...

Transfer learning has become a common practice for training deep models with limited labeled data in target domain. On the other hand, are vulnerable to adversarial attacks. Though transfer been widely applied, its effect on model robustness is unclear. To figure out this problem, we conduct extensive empirical evaluations show that fine-tuning effectively enhances under white-box FGSM We also propose black-box attack method which attacks examples produced by source model. systematically...

Greedy Equivalent Search (GES) is an effective algorithm for Bayesian network problem, which searches in the space of graph equivalence classes. However, original GES may easily fall into local optimization trap because empty initial structure. In this paper, An improved method prosposed. It firstly makes a draft real network, based on Maximum Information Coefficient (MIC) and conditional independence tests. After step, many independent relations can be found. To ensure correctness, then...

Topic models are widely explored for summarizing a corpus of documents. Recent advances in Variational AutoEncoder (VAE) have enabled the development black-box inference methods topic modeling order to alleviate drawbacks classical statistical inference. Most existing VAE based approaches assume unimodal Gaussian distribution approximate posterior latent variables, which limits flexibility encoding space. In addition, unsupervised architecture hinders incorporation extra label information,...

Parameters in deep neural networks which are trained on large-scale databases can generalize across multiple domains, is referred as "transferability". Unfortunately, the transferability usually defined discrete states and it differs with domains network architectures. Existing works heuristically apply parameter-sharing or fine-tuning, there no principled approach to learn a parameter transfer strategy. To address gap, unit (PTU) proposed this paper. The PTU learns fine-grained nonlinear...

In this paper we propose a novel multiple target tracking model composed of two detectors and tracker. An on-line detector tracker are used to generate candidates, whose confidence scores then evaluated by the off-line trained detectors. data association stage, high-efficient inference in structural leads optimal result. The experimental results demonstrate that our can overcome occlusion appearance changing problems. be applied analyze information single or among targets.

Extracting rich feature representations is a key challenge in person re-identification (Re-ID) tasks. However, traditional Convolutional Neural Networks (CNN) based methods could ignore part of information when processing local regions images, which leads to incomplete extraction. To this end, paper proposes Re-ID method on vision Transformer with hierarchical structure and window shifting. When extracting image features, the model constructed by introducing construction commonly used CNN....

Pedestrian detection is one of the key technologies in automotive safety, robotic and intelligent video surveillance. Recently, deep convolutional neural networks have achieved significant effect image classification retrieval tasks. In this paper, we propose a novel model for pedestrian to simultaneously extract classify features. The proposed 19 layers network which consists 7 convolution layers, 3 pooling 6 relu 2 normalization layer. classical back propagation algorithm adopted train...

Fine-tuning can be vulnerable to adversarial attacks. Existing works about black-box attacks on fine-tuned models (BAFT) are limited by strong assumptions. To fill the gap, we propose two novel BAFT settings, cross-domain and cross-architecture BAFT, which only assume that (1) target model for attacking is a model, (2) source domain data known accessible. successfully attack under both first train an generator against adopts encoder-decoder architecture maps clean input example. Then search...

Deep domain adaptation models learn a neural network in an unlabeled target by leveraging the knowledge from labeled source domain. This can be achieved learning domain-invariant feature space. Though learned representations are separable domain, they usually have large variance and samples with different class labels tend to overlap which yields suboptimal performance. To fill gap, Fisher loss is proposed discriminative within-class compact between-class separable. Experimental results on...

The various scale makes detecting and localizing objects a challenging problem, especially for small-scale instances [1, 2]. While most existing models focus on detection in static images, we investigate the video surveillance scenario. In this paper, probabilistic graphical model is proposed to integrate local generic object detector scene-specific contextual features. outperforms part-based by extending them into multiresolution structure. Experimental results public dataset CAVIAR [3]...

The paper develops a connection between the first-order logic representation and content structure model in sentiment analysis applications. We propose modified semi-supervised approach to study word-level with well-designed features. is Conditional Random Fields (CRF) latent topic nodes. Introducing features into our can solve long-distance dependency problem. new applied two multi-aspect tasks: sentence labeling task rating prediction task. use data from Amazon corpus movie-review corpus....

Finding the optimistic triangulation in Bayesian network, is NP hard. Optimization Algorithm a new kind of evolutionary algorithm estimation distribution algorithms (EDAs). An improved BOA proposed to get approximate this paper. We carry out four EDAs including our method, on standard networks. Comparing with other Estimation algorithms, method displays better experimental results and robustness.

The self-organizing map (SOM) is an excellent tool for data mining. In order to know the understanding of businesses establishing environment-friendly enterprise, we carried out a questionnaire 49 manufacturing enterprises in city and clustered results using SOM network. We found that there are four different classes among enterprises. doing this work developed new procedure uniting one-dimensional with two-dimensional visualizing exploring properties training results. can reduce subjective...