Search

Graph neural networks (GNNs) are gaining increasing popularity as a promising approach to machine learning on graphs. Unlike traditional graph workloads where each vertex/edge is associated with a scalar, GNNs attach a feature tensor to each vertex/edge. This additional feature dimension, along with consequently more complex vertex- and edge-wise computations, has enormous implications on locality and parallelism, which existing graph processing systems fail to exploit. This paper proposes FeatGraph to accelerate GNN workloads by co-optimizing graph traversal and feature dimension computation. FeatGraph provides a flexible programming interface to express diverse GNN models by composing coarse-grained sparse templates with fine-grained user-defined functions (UDFs) on each vertex/edge. FeatGraph incorporates optimizations for graph traversal into the sparse templates and allows users to specify optimizations for UDFs with a feature dimension schedule (FDS). FeatGraph speeds up end-to-end GNN training and inference by up to 32x on CPU and 7x on GPU., Comment: SC'20; changed all figures to type 1

It has been found that the research of a special kind of frames, named $K$ -frames where $K$ is an operator, is significant in theory and application. In this paper, first of all, in theory, we discuss some properties about the existence and structure of the $K$ -frames and the relationship between operator $K$ and $K$ -frames. And these properties provide a method to construct $K$ -frames in the following text. Next, in application, in order to use $K$ -frames to deal with the data erasures of communication, we establish a probability model. Based on this probability model and the above-mentioned properties of $K$ -frames, we construct a probability uniform Parseval $K$ -frame, such that its canonical dual is the unique probability optimal dual when $\|Kf\|\setminus \|f\| =c$ and $\|K^{+}f\|\setminus \|f\| =a$ , where both $a$ and $c$ are constants, where $K^{+}$ is the pseudoinverse for $K$ . We use $\|f\|$ to represent the norm of $f$ , where $f$ is any vector in Hilbert space. Moreover, the existence of the probability uniform Parseval $K$ -frame is discussed and then uses it to handle the problem of data erasures. Furthermore, we obtain the conditions for operator $K$ to form the $K$ -frame, which lead to the reduction of restrictions on frame sequences, and make the construction become more flexible compared with traditional frames in coding. Finally, we use three figures obtained by numerical experiments to compare the decoding effects.

Brain-computer interface (BCI) instead of depending on the brain's normal output pathways, can use electroencephalogram (EEG) from the scalp as the representation of brain activity to control external devices. EEG during motor imagery (MI) provides a non-muscular communication way to control external devices and has advantage of non-invasiveness and high time resolution. However the application is still limited by time-consuming training and poor classification rate with multiple categories etc. We recorded 64-channel scalp EEG from eight healthy subjects during imagery tasks of left, right hand movements and stop. EEG was analyzed in time-frequency distribution and spatial topographies were explored too. A one versus one common spatial pattern was applied to construct feature vector and then linear discriminant analysis was used for the classification. For the purpose of real time control in the future, small training size was used and we got discrimination among three types of motor imagery at the accuracy rate about 90%.

To develop prosthesis controlled by spontaneous brain signal, we studied the motion intention underlying EEG activity with the imagination of right hand and left hand. During hand motor imaginary, EEG activities are collected by 64 electrodes from healthy adults. The temporal and frequency characteristics associated with the imagination of movement are explored and combined as the feature vector for the linear discriminant analysis. Our findings show a classification accuracy of 77% for our datasets and 83% for the datasets of BCI Competition 2003.

Intelligent prosthetics aim to provide a communication channel between disabled people and external world. Among them, walking-assistive device to help people with lower-limb disability attracts attention but still limited by unnatural gait or lack accurate real-time control. Here we studied the coupling between the upper limbs and the lower limbs since the whole body actually involved during walking and there exists coordination among limbs. We collected motion information of different joints from healthy people during bipedal walking by attitudes sensors. After gait cycle identified and normalized, the relationship between different limbs was built. Based on the polynomial approximation, the knee joint angle can be estimated by the shoulder. Thus in the future it can be potentially applied to control walking-assistive prosthetics by the upper limbs movement instead of the motion information of lower limbs.

A continuous attractor of network is a set of connected stable equilibrium points and it has been recognized as a suitable model for describing the encoding of continuous stimuli in neural networks. Population coding is a method to represent stimuli by using the collective activities of a number of neurons. It is difficult to extract information from these population codes because of the noise inherent in neuronal responses. In this paper the dynamics analysis of a population decoding model is investigated. The conditions of the population decoding model to possess nonzero continuous attractor are derived. Moreover the explicit expressions of the continuous attractors are provided. Simulations are employed to illustrate the theory.

As a kind of statistical features, texture features often have a rotary deformation, and have strong resistibility to noise. The paper first constructs the gray level co-occurrence matrix of face image to describe texture feature of face image, and then uses the classification method of minimum weighted Euclidean distance to fulfill the matching and identification of face. Experiments results have shown that recognition rate was greatly increased by the combination of weighted Euclidean distance and texture feature.

The dynamics of a class of recurrent neural networks with uniform constant external input is investigated in this paper. Nowadays, continuous attractors have been recognized as promising model for describing the encoding of continuous stimuli in neural networks. The existence of the continuous attractors depends on a lot of factors such as connection wight and the external input and so on. The external input is very important in allowing the network to exhibit such an interesting character. Small changes in this external input level may shift a network from a relatively quiet state to some other state with highly complex dynamics. Thus, the external input may act as a switch that allows networks to be turned on or off regardless of their function. In this paper we find when the external input changes from zero to a small positive value, the steady state of network will shift from one continuous attractor to another. It shows how the external input controls the state of networks. Examples and simulation results are used to illustrate the theory.

Wu Cun Shan Tunnel in Xiamen Airport Road Phase I Project is a shallow-buried, large-span and hidden-digging tunnel which passes through high-density building cluster under complicated geological conditions. 67 buildings are being influenced by tunnel construction along the tunnel. In order to ensure the safety of surface buildings in the tunnel construction, CRD excavation method are adopted in some sections which initiates a precedent of construction of double-arch tunnel with CRD method in China. The effect of tunnel construction on frame structure buildings is analyzed through site monitoring of surface building settlement, crack development in this paper. The results show that excavation of tunnel face CRD1∼5 has the most important influence on the building settlement. The longer the construction period is, the greater building settlement and crack width change, and vice versa. Grouting in the tunnel can decrease building settlement caused by tunnel excavation effectively. The main reason that causes building cracking and development of cracks existed is uneven settlement. And the ability of the building to resist down concave deformation is stronger than that to upper convex deformation. While width variation of cracks lags behind variation of uneven settlement. Cracks caused by tunnel construction mainly distribute around the door and window openings in the longitudinal filler wall and the connection part between frame beam-column and filler wall.

Researches on neural population coding have revealed that continuous stimuli, such as orientation, moving direction, and the spatial location of objects could be encoded as continuous attractors in neural networks. The dynamical behaviors of continuous attractors are interesting properties of recurrent neural networks. This paper proposes a class of recurrent neural networks without lateral inhibition. Since there is no general rule to determine the stability of the network without specifying the excitatory connections, individual conditions can be calculated analytically for some particular cases. It shows that the networks can possess continuous attractors if the excitatory connections are in gaussian shape. Simulation examples are employed for illustration.

This paper proposes a new approach to solve Traveling Salesman Problems (TSPs) by using a class of Lotka-Volterra neural networks (LVNN) without self-excitatory. Some stability criteria that ensure the convergence of valid solutions are obtained. It is proved that a class of equilibrium states are stable if and only if they correspond to the valid solutions of the TSPs. That is, one can always obtain a valid solution whenever the network convergence to a stable state. A set of analytical conditions for optimal settings of LVNN is derived. The simulation results illustrate the theoretical analysis.

To describe the encoding of continuous stimuli in neural networks, continuous attractors have been recognized as promising models. A continuous attractor is a set of connected stable equilibrium points. It exhibits interesting dynamical properties in many recurrent neural networks. This paper studies the continuous attractors of discrete-time cellular neural networks (DCNNs). The main contribution is that the representations of continuous attractors for DCNNs are obtained under some conditions. Such important results provide clear and complete descriptions to the continuous attractors of DCNNs.