Abstract While existing supervised semantic segmentation methods have shown significant performance improvements, they heavily rely on large‐scale pixel‐level annotated data. To reduce this dependence, recent research has proposed semi‐supervised learning‐based methods that have achieved great success. However, almost all these works are mainly dedicated to image semantic segmentation, while semi‐supervised video semantic segmentation (SVSS) has been barely explored. Due to the significant difference between video data and image, simply adapting semi‐supervised image semantic segmentation approaches to SVSS may neglect the inherent temporal correlations in video frames. This paper presents a novel method (named TRACL) with temporal reconstruction (TR) and adaptive consistency loss (ACL) for SVSS, aiming to fully utilize the temporal relations of internal frames in video clip. The authors’ TR method implements the reconstruction from the feature and output levels to narrow the distribution gap between internal video frames. Specifically, considering the underlying data distribution, the authors construct Gaussian models for each category, and use probability density function to obtain the similarity between different feature maps for temporal feature reconstruction. The authors’ ACL can adaptively select two pixel‐wise consistency loss including Flow Consistency Loss and Reconstruction Consistency Loss, providing stronger supervision signals for unlabelled frames during model training. Additionally, the authors extend their method to unlabelled video for more training data by employing mean‐teacher structure. Extensive experiments on three datasets including Cityscapes, Camvid and VSPW demonstrate that the authors’ proposed method outperforms previous state‐of‐the‐art methods.

In black-box scenarios, most transfer-based attacks usually improve the transferability of adversarial examples by optimizing the gradient calculation of the input image. Unfortunately, since the gradient information is only calculated and optimized for each pixel point in the image individually, the generated adversarial examples tend to overfit the local model and have poor transferability to the target model. To tackle the issue, we propose a resize-invariant method (RIM) and a logical ensemble transformation method (LETM) to enhance the transferability of adversarial examples. Specifically, RIM is inspired by the resize-invariant property of Deep Neural Networks (DNNs). The range of resizable pixel is first divided into multiple intervals, and then the input image is randomly resized and padded within each interval. Finally, LETM performs logical ensemble of multiple images after RIM transformation to calculate the final gradient update direction. The proposed method adequately considers the information of each pixel in the image and the surrounding pixels. The probability of duplication of image transformations is minimized and the overfitting effect of adversarial examples is effectively mitigated. Numerous experiments on the ImageNet dataset show that our approach outperforms other advanced methods and is capable of generating more transferable adversarial examples.

Variational dropout (VD) is a generalization of Gaussian dropout, which aims at inferring the posterior of network weights based on a log-uniform prior on them to learn these weights as well as dropout rate simultaneously. The log-uniform prior not only interprets the regularization capacity of Gaussian dropout in network training, but also underpins the inference of such posterior. However, the log-uniform prior is an improper prior (i.e., its integral is infinite), which causes the inference of posterior to be ill-posed, thus restricting the regularization performance of VD. To address this problem, we present a new generalization of Gaussian dropout, termed variational Bayesian dropout (VBD), which turns to exploit a hierarchical prior on the network weights and infer a new joint posterior. Specifically, we implement the hierarchical prior as a zero-mean Gaussian distribution with variance sampled from a uniform hyper-prior. Then, we incorporate such a prior into inferring the joint posterior over network weights and the variance in the hierarchical prior, with which both the network training and dropout rate estimation can be cast into a joint optimization problem. More importantly, the hierarchical prior is a proper prior which enables the inference of posterior to be well-posed. In addition, we further show that the proposed VBD can be seamlessly applied to network compression. Experiments on classification and network compression demonstrate the superior performance of the proposed VBD in regularizing network training.

Non-negative matrix factorization (NMF) is a challenging problem due to its ill-posed nature. The key for the success of NMF is to exploit appropriate prior models for those two decomposed factor matrices. Although lots of effective sparsity-inducing prior models have been developed for NMF, they are often rooted in either ℓpregularization with p > 0, which only provide an approximation to the ℓ0sparsity, ultimately resulting in a sub-optimal solution. To address this problem, we propose a novel truncated spike-and-slab prior based Bayesian NMF method. Through integrating a Bernoulli distribution with a truncated Gaussian distribution together, the proposed prior is capable of imposing the exact ℓ0regularization as well as the non-negativity constraint on the factor matrices. Further, the proposed prior can be extended to robust NMF problem. Experimental results in blind source separation, face images representation and image denoising demonstrate the advantage of the proposed method.

Variational Bayesian (VB) learning has been successfully applied to instantaneous blind source separation. However, the traditional VB learning is restricted to the separation of independent source signals. Moreover, it has the difficulty to recover source signals with a sizable number of samples because of its rapidly increasing computational requirement. To overcome such shortcomings, frame-based VB (FVB) learning is proposed to address both independent and dependent source separation with a large number of samples in this paper. Specifically, a Gaussian process (GP) is employed to model independent or dependent source signals. To our knowledge, GP has been only used to model each of independent source signals. For dependent source signals, this paper proposes a novel modeling process: initial source signals are zigzag concatenated into a long serial and GP is then used to model it. In order to obtain a reliable covariance function for GP, first, we apply singular value decomposition to give initial estimated source signals and then we select an appropriate covariance function with which GP can perfectly fit them. In order to alleviate the computational burden of VB learning, we split observed signals into frames, and then model and infer source signals for each frame. Compared with the state-of-the-art algorithms, the experimental results show that the FVB learning has potential to provide improvement in separation performance not only for independent source signals but also for dependent ones, especially for long data records.

This paper presents an adaptive particle swarm optimization with supervised learning and control (APSO-SLC) for the parameter settings and diversity maintenance of particle swarm optimization (PSO) to adaptively choose parameters, while improving its exploration competence. Although PSO is a powerful optimization method, it faces such issues as difficult parameter setting and premature convergence. Inspired by supervised learning and predictive control strategies from machine learning and control fields, we propose APSO-SLC that employs several strategies to address these issues. First, we treat PSO with its optimization problem as a system to be controlled and model it as a dynamic quadratic programming model with box constraints. Its parameters are estimated by the recursive least squares with a dynamic forgetting factor to enhance better parameter setting and weaken worse ones. Its optimal parameters are calculated by this model to feed back to PSO. Second, a progress vector is proposed to monitor the progress rate for judging whether premature convergence happens. By studying the reason of premature convergence, this work proposes the strategies of back diffusion and new attractor learning to extend swam diversity, and speed up the convergence. Experiments are performed on many benchmark functions to compare APSO-SLC with the state-of-the-art PSOs. The results show that it is simple to program and understand, and can provide excellent and consistent performance.

In order to improve the performance of PSO, this paper presents an Autonomous Learning Adaptation method for Particle Swarm Optimization (ALA-PSO) to automatically tune the control parameters of each particle. Although PSO is an ideal optimizer, one of its drawbacks focuses on its performance dependency on its parameters, which differ from one problem to another. In ALA-PSO, each particle is viewed as an intelligent agent and aims at improving itself performance, and can autonomously learn how to tune its parameters from its own experiment of successes and failures. For each particle, it means successful movement if the value of objective function in current position is improved than previous position, otherwise means failure. In case of successful movement, the parameters that are positive correlation with the direction of forward movement should be increased otherwise should be decreased. Meanwhile, in case of unsuccessful movement, inverse operation should be performed. The proposed parameter adaptive method is compared with several existing adaptive strategies, and the results show that ALA-PSO is not only effective, but also robust in different categories benchmarks.

To improve the efficiency of a face detector, this paper presents an automatic distributing detector (ADD) based on the fuzzy theory to improve the performance of face detection. The main contributions lie in:l) A new Haar-like feature representation based on the fuzzy membership function is proposed, 2)The entropy of feature set is employed as choice criteria to select weak classifiers, 3) The AdaBoost algorithm is used to train weak classifiers, and 4)The distributor which can dynamically select stronger classifiers is constructed. The experiment results show that the proposed method not only determines rapidly the sub-window which contains the human face, but also tune the classifier dynamically to adaptive new samples. The accuracy and speed of our method are also promoted comparison with the state-of-art detectors. On the other hand, as for the image sub-window which is like face, according to the value of membership function, distributor can dynamically select the remaining stronger classifiers to determine. This detector can effectually improve detection speed and has better detection performance.

This paper presents a Gaussian classifier-based evolutionary strategy (GCES) to solve multimodal optimization problems. An evolutionary technique for them must answer two crucial questions to guarantee its success: how to distinguish among the different basins of attraction and how to safeguard the already discovered good-quality solutions including both global and local optima. In GCES, multimodal optimization problems are regarded as classification ones, and Gaussian mixture models are employed to save the locations and basins of already and presently identified local or global optima. A sequential estimation technique for the covariance of a Gaussian model is introduced into GCES. To best adjust the global step size, a strategy named top-ranked sample selection is introduced, and a classification method instead of a common but problematic radius-triggered manner is proposed. Experiments are performed on a series of benchmark test functions to compare GCES with the state-of-the-art multimodal optimization approaches. The results show that GCES is not only simple to program and understand, but also provides better and consistent performance.

One of the main unsolved problems in computer algebra is to determine the minimal number of multiplications which is necessary to compute the product of two matrices. For practical value, the small format is of special interest. This leads to a combinatorial optimization problem which is unlikely solved in polynomial time. In this paper, we present a method called combining Gaussian eliminations to reduce the number of variables in this optimization problem and use heuristic ant colony algorithm to solve the problem. The results of experiments on 2 × 2 case show that our algorithm achieves significant performance gains. Extending this algorithm from 2 × 2 case to 3 × 3 case is also discussed. Index Terms—Ant colony optimization (ACO), evolutionary algorithms, Gaussian eliminations, matrix multiplication, multiplicative complexity, Strassen's algorithm.