Research on a Convolution Kernel Initialization Method for Speeding Up the Convergence of CNN

This paper presents a convolution kernel initialization method based on the local binary patterns (LBP) algorithm and sparse autoencoder. This method can be applied to the initialization of the convolution kernel in the convolutional neural network (CNN). The main function of the convolution kernel is to extract the local pattern of the image by template matching as the target feature of subsequent image recognition. In general, the Xavier initialization method and the He initialization method are used to initialize the convolution kernel. In this paper, firstly, some typical sample images were selected from the training set, and the LBP algorithm was applied to extract the texture information of the typical sample images. Then, the texture information was divided into several small blocks, and these blocks were input into the sparse autoencoder (SAE) for pre-training. After finishing the training, the weight values of the sparse autoencoder that met the statistical features of the data set were used as the initial value of the convolution kernel in the CNN. The experimental result indicates that the method proposed in this paper can speed up the convergence of the network in the network training process and improve the recognition rate of the network to an extent.

Pattern similarity in the frontoparietal control network reflects an "off-veridical" template that optimizes target-match decisions during visual search

Theories of attention hypothesize the existence of an "attentional" or "target" template that contains task-relevant information in memory when searching for an object. The target template contributes to visual search by directing visual attention towards potential targets and serving as a decisional boundary for target identification. However, debate still exists regarding how template information is stored in the human brain. Here, we conducted a pattern-based fMRI study to assess how template information is encoded to optimize target-match decisions during visual search. To ensure that match decisions reflect visual search demands, we used a visual search paradigm in which all distractors were linearly separable but highly similar to the target and were known to shift the target representation away from the distractor features (Yu & Geng, 2019). In a separate match-to-sample probe task, we measured the target representation used for match decisions across two resting state networks that have long been hypothesized to maintain and control target information: the frontoparietal control network (FPCN) and the visual network (VisN). Our results showed that lateral prefrontal cortex in FPCN maintained the context-dependent "off-veridical" template; in contrast, VisN encoded a veridical copy of the target feature during match decisions. By using behavioral drift diffusion modeling, we verified that the decision criterion during visual search and the probe task relied on a common biased target template. Taken together, our results suggest that sensory-veridical information is transformed in lateral prefrontal cortex into an adaptive code of target-relevant information that optimizes decision processes during visual search.

Research on Target Tracking Algorithm of Twin Networks Integrating Attention Mechanism

Aiming at the current stage of the twin network target tracking algorithm, the tracking target is occluded, the tracking is affected by illumination, and the target's scale change from far to near or from near to far causes tracking failure. This article will optimize and improve from two directions. The twin neural network first uses an adaptive detailed feature extraction, adds a residual network to the twin network, and embeds a detailed feature retention module in each layer, amplifies the changes in the target feature, and retains the important structure of the original target feature Details: Secondly, the introduction of a spatial attention mechanism allows the main branch to pay more attention to the area to be matched, improves the ability to distinguish features, and makes the tracking effect better. In order to verify the effectiveness of this experiment, this experiment was tested on the data set OTB2015. The experiment proved that the proposed algorithm performs better in accuracy and success rate.

Sound Event Detection by Pseudo-Labeling in Weakly Labeled Dataset

Weakly labeled sound event detection (WSED) is an important task as it can facilitate the data collection efforts before constructing a strongly labeled sound event dataset. Recent high performance in deep learning-based WSED’s exploited using a segmentation mask for detecting the target feature map. However, achieving accurate detection performance was limited in real streaming audio due to the following reasons. First, the convolutional neural networks (CNN) employed in the segmentation mask extraction process do not appropriately highlight the importance of feature as the feature is extracted without pooling operations, and, concurrently, a small size kernel forces the receptive field small, making it difficult to learn various patterns. Second, as feature maps are obtained in an end-to-end fashion, the WSED model would be weak to unknown contents in the wild. These limitations would lead to generating undesired feature maps, such as noise in the unseen environment. This paper addresses these issues by constructing a more efficient model by employing a gated linear unit (GLU) and dilated convolution to improve the problems of de-emphasizing importance and lack of receptive field. In addition, this paper proposes pseudo-label-based learning for classifying target contents and unknown contents by adding ’noise label’ and ’noise loss’ so that unknown contents can be separated as much as possible through the noise label. The experiment is performed by mixing DCASE 2018 task1 acoustic scene data and task2 sound event data. The experimental results show that the proposed SED model achieves the best F1 performance with 59.7% at 0 SNR, 64.5% at 10 SNR, and 65.9% at 20 SNR. These results represent an improvement of 17.7%, 16.9%, and 16.5%, respectively, over the baseline.

BRICS Network University potential as an educational platform for innovative cooperation

The prospects for increasing the share of training and coverage of educational cooperation in the field of innovation in the BRICS Network University - one of the most important modern forms of cooperation in the field of education at the international level - are studied. The concepts of network structure and network university are considered. The effects of network interaction and the principles of functioning of network universities are revealed, and an overview of the approach to classifying network structures by the target feature of formation is given. The purpose is to search for opportunities to set up an innovative orientation of the BRICS Network University educational process as the general innovative cooperation stimulation platform of the participating countries. For this purpose, the experience of participation of Russian universities in other network universities (CIS and SCO) was explored, and possible formats of cooperation between innovation teams and interested students in the BRICS countries were proposed to stimulate both traditional mobility of students, teachers and employees, and general innovation exchange through the BRICS Network University.

Multi-modal Data Fusion Method for Human Behavior Recognition Based on Two IA-Net and CHMM

The multi-modal data fusion method based on IA-net and CHMM technical proposed is designed to solve the problem that the incompleteness of target behavior information in complex family environment leads to the low accuracy of human behavior recognition.The two improved neural networks(STA-ResNet50、STA-GoogleNet)are combined with LSTM to form two IA-Nets respectively to extract RGB and skeleton modal behavior features in video. The two modal feature sequences are input CHMM to construct the probability fusion model of multi-modal behavior recognition.The experimental results show that the human behavior recognition model proposed in this paper has higher accuracy than the previous fusion methods on HMDB51 and UCF101 datasets. New contributions: attention mechanism is introduced to improve the efficiency of video target feature extraction and utilization. A skeleton based feature extraction framework is proposed, which can be used for human behavior recognition in complex environment. In the field of human behavior recognition, probability theory and neural network are cleverly combined and applied, which provides a new method for multi-modal information fusion.

Which animal is big? Aging-related effects on the controlled retrieval of information

The efficient use of knowledge requires semantic control processes to retrieve context-relevant information. So far, it is well established that semantic knowledge, as measured with vocabulary tests, do not decline in aging. Yet, it is still unclear if controlled retrieval -the context-driven retrieval of very specific aspects of semantic knowledge- declines in aging, following the same fate of other forms of cognitive control. Here, we tackled this issue by comparing the performance of younger and older native Italian speakers during a semantic feature verification task. To manipulate the control demands, we parametrically varied the semantic significance, a measure of the salience of the target feature for the cue concept. As compared to their young counterparts, older adults showed a greater performance disruption (in terms of reaction times) as the significance value of the target feature decreased. This result suggests that older people have difficulties in regulating the activation within semantic representation, such that they fail to handle non-dominant (or weakly activated) yet task-relevant semantic information.

Feature distribution learning by passive exposure

Humans can rapidly estimate the statistical properties of groups of stimuli, including their average and variability. But recent studies of so-called Feature Distribution Learning (FDL) have shown that observers can quickly learn even more complex aspects of feature distributions. In FDL, observers learn the full shape of a distribution of features in a set of distractor stimuli and use this information to improve visual search: response times (RT) are slowed if the target feature lies inside the previous distractor distribution, and the RT patterns closely reflect the distribution shape. FDL requires only a few trials and is markedly sensitive to different distribution types. It is unknown, however, whether our perceptual system encodes feature distributions automatically and by passive exposure, or whether this learning requires active engagement with the stimuli. In two experiments, we sought to answer this question. During an initial exposure stage, participants passively viewed a display of 36 lines that included one orientation singleton or no singletons. In the following search display, they had to find an oddly oriented target. The orientations of the lines were determined either by a Gaussian or a uniform distribution. We found evidence for FDL only when the passive trials contained an orientation singleton. Under these conditions, RT decreased as a function of the orientation distance between the target and the exposed distractor distribution. These results suggest that FDL can occur by passive exposure, but only if an orientation singleton appears during exposure to the distribution.