Search

Occlusion problem remains a key challenge in Optical Flow Estimation (OFE) despite the recent significant progress brought by deep learning in the field. Most existing deep learning OFE methods, especially those based on two frames, cannot properly handle occlusions, in part because there is no significant feature similarity in occluded regions. The multi-frame settings have the potential to mitigate the occlusion issue in OFE. However, the problem of Multi-frame OFE (MOFE) remains underexplored, and the limited works are specially designed for pyramid backbones and obtain the aligned temporal information by time-consuming backward flow calculation or non-differentiable forward warping transformation. To address these shortcomings, we propose an efficient MOFE framework named SplatFlow, which is realized by introducing the differentiable splatting transformation to align the temporal information, designing a One-to-Many embedding method to densely guide the current frame's estimation, and further remodelling the existing two-frame backbones. The proposed SplatFlow is very efficient yet more accurate as it is able to handle occlusions properly. Extensive experimental evaluations show that our SplatFlow substantially outperforms all published methods on KITTI2015 and Sintel benchmarks. Especially on Sintel benchmark, SplatFlow achieves errors of 1.12 (clean pass) and 2.07 (final pass), with surprisingly significant 19.4% and 16.2% error reductions from the previous best results submitted, respectively. Code is available at https://github.com/wwsource/SplatFlow.

Binary neural networks (BNNs), where both weights and activations are binarized into 1 bit, have been widely studied in recent years due to its great benefit of highly accelerated computation and substantially reduced memory footprint that appeal to the development of resource constrained devices. In contrast to previous methods tending to reduce the quantization error for training BNN structures, we argue that the binarized convolution process owns an increasing linearity towards the target of minimizing such error, which in turn hampers BNN's discriminative ability. In this paper, we re-investigate and tune proper non-linear modules to fix that contradiction, leading to a strong baseline which achieves state-of-the-art performance on the large-scale ImageNet dataset in terms of accuracy and training efficiency. To go further, we find that the proposed BNN model still has much potential to be compressed by making a better use of the efficient binary operations, without losing accuracy. In addition, the limited capacity of the BNN model can also be increased with the help of group execution. Based on these insights, we are able to improve the baseline with an additional 4~5% top-1 accuracy gain even with less computational cost. Our code will be made public at https://github.com/zhuogege1943/ftbnn., Comment: Openreview: https://openreview.net/forum?id=9wHe4F-lpp

In the paper, as an improvement of fuzzy clustering neural network FCNN proposed by Zhang et al., a novel robust fuzzy clustering neural network RFCNN is presented to cope with the sensitive issue of clustering when outliers exist. This new algorithm is based on Vapnik's @?-insensitive loss function and quadratic programming optimization. Our experimental results demonstrate that RFCNN has much better robustness for outliers than FCNN.

In this paper, a novel robust maximum entropy clustering algorithm RMEC, as the improved version of the maximum entropy algorithm MEC [2–4], is presented to overcome MEC's drawbacks: very sensitive to outliers and uneasy to label them. Algorithm RMEC incorporates Vapnik's ɛ – insensitive loss function and the new concept of weight factors into its objective function and consequently, its new update rules are derived according to the Lagrangian optimization theory. Compared with algorithm MEC, the main contributions of algorithm RMEC exit in its much better robustness to outliers and the fact that it can effectively label outliers in the dataset using the obtained weight factors. Our experimental results demonstrate its superior performance in enhancing the robustness and labeling outliers in the dataset.

In [1], with the evidence framework, the almost inversely linear dependency between the optimal parameter r in norm-r support vector regression machine r-SVR and the Gaussian input noise is theoretically derived. When r takes a non-integer value, r-SVR cannot be easily realized using the classical QP optimization method. This correspondence attempts to achieve two goals: (1) The Newton-decent-method based implementation procedure of r-SVR is presented here; (2) With this procedure, the experimental studies on the dependency between the optimal parameter r in r-SVR and the Gaussian noisy input are given. Our experimental results here confirm the theoretical claim in [1].

When applying fuzzy systems for data analysis, their approximation and interpretation capabilities are two important aspects. Cascaded fuzzy system (CFS) is a new special class of hierarchical fuzzy systems in architectures proposed by Duan and Chung [IEEE Trans. Fuzzy Syst. 9 (2) (2001) 293] but its universal approximation capability is still not proved. When CFS is utilized in fuzzy data analysis/modeling, it seems very difficult to give a reasonable interpretation for intermediate variables and the corresponding fuzzy rules. A new cascaded centralized TSK fuzzy system (CCTSKFS) is presented in this paper, whose universal approximation capability is proved in detail, and what's more, we can interpret CCTSKFS more rationally. Finally, our experimental results demonstrate that CCTSKFS outperforms the classical cascaded TSK fuzzy system (CTSKFS) in approximation capability and robustness.

In this paper the novel fuzzy kernel hyperball perceptron is presented. The proposed method first maps the input data into a high-dimensional feature space using some Mercer kernel function. Then, the decision function for each class is derived by the learning rules of the fuzzy kernel hyperball perceptron. The fuzzy membership functions, which resolve unclassifiable zones among classes, are incorporated into its classification algorithm to further enhance the perceptron’s adaptability and classification accuracy effectively. Unlike SVM, the fuzzy kernel hyperball perceptron has no convergence problem and avoids solving so-called quadratic programming problem which often makes SVM ineffective for large data sets. Especially, unlike the classical SVMs, we can directly utilize the fuzzy kernel hyperball perceptrons to solve multiclass problems, without using any pairwise combination. Our experimental results demonstrate its effectiveness.

With the evidence framework, the regularized linear regression model can be explained as the corresponding MAP problem in this paper, and the general dependency relationships that the optimal parameters in this model with noisy input should follow is then derived. The support vector regression machines Huber-SVR and Norm-r r-SVR are two typical examples of this model and their optimal parameter choices are paid particular attention. It turns out that with the existence of the typical Gaussian noisy input, the parameter μ in Huber-SVR has the linear dependency with the input noise, and the parameter r in the r-SVR has the inversely proportional to the input noise. The theoretical results here will be helpful for us to apply kernel-based regression techniques effectively in practical applications.

Proceedings of the 6th International Brain-Computer Interface Conference 2014

In this paper the four-dimensional dynamic diffusing mechanism and the enhancement in Long-Term Potentiation (LTP) of intrinsic nitric oxide (NO) in nervous system are studied computationally. A novel unsupervised Diffusing Self-Organizing Maps (DSOM) model is presented on the union of SOM with NO diffusing mechanism. Based on the spatial prototype mapping, temporal enhancement is introduced in DSOM and the fine-tuning manner is improved by the simplified NO diffusing mechanism. Furthermore, the quantization error of optimal weights is valuated and the detailed noise analysis of DSOM is presented. Finally some typical stimulation experiments are presented to illustrate how DSOM gracefully handles time warping and multiple patterns with overlapping reference vectors.