Search

Memristor based neural networks have great potentials in on-chip neuromorphic computing systems due to their properties of fast computation and low-energy consumption. However, the imprecise features of synaptic weight encoding by memristor devices generally result in catastrophic failures of the network in situ training. Learning schemes that consider all of the imprecise memristor network properties were rarely reported, such as variations from device-to-device (D2D) and pulse-to-pulse (P2P), weight programming without reading, pulse number rounding, etc. In this work, we have considered all major imprecise proprieties and designed a learning scheme that integrates stochastic sparse updating with momentum adaption (SSM) to efficiently train the imprecise memristor networks with high classification accuracy. With the SSM scheme, experiments show that the classification accuracy on multilayer perceptron (MLP) and convolutional neural network (CNN) improves from 26.12% to 90.07% and from 65.98% to 92.38%, respectively. Meanwhile, the total numbers of weight updating pulses decrease 90% and 40% in MLP and CNN, respectively, and the convergence rates are both 3 × faster. The SSM scheme provides a high-accuracy, low-power, and fast-convergence solution for the in situ training of imprecise memristor networks, which is crucial to future neuromorphic intelligence systems.

Target control of complex networks, which aims to control only a target subset of network nodes instead of the entire network, is an outstanding challenge faced in various real world applications. Recently one fundamental issue regarding how to allocate a minimum number of control sources to guarantee the target controllability of a given target node set S was solved. This issue is shown to be essentially a path cover problem and it can be converted to a maximum network flow problem. In this work, we address another fundamental issue which is to further find the maximum (minimum) controllable node set to cover S using directed paths and circles based on the allocated minimum number of control sources. We show that such an issue can be solved by applying the maximum (minimum) cost maximum flow algorithm after introducing a cost for each edge in the flow network. Based on the obtained maximum (minimum) target controllable node set, we further propose a new index termed “expandable target controllability” to characterize complex networks that are expanding all the time. It is shown that “expandable target controllability” is an intrinsic property of various networks. We anticipate that this work would serve wide applications in target control and expandable target control of real-life networks.

Super-resolution (SR) is a useful technology to generate a high-resolution (HR) visual output from the low-resolution (LR) visual inputs overcoming the physical limitations of the cameras. However, SR has not been applied to enhance the spatial resolution of event-stream images captured by the frame-free dynamic vision sensor (DVS). SR of event-stream image aims to recover the same statistics of events which is fundamentally different from the existing frame-based schemes. In this work, a two-stage scheme is proposed to solve the spatial SR problem of the spatiotemporal event-stream image. We use a nonhomogeneous Poisson point process to model the event sequence, and sample the events of each pixel by simulating a nonhomogeneous Poisson process according to the specified event number and rate function. Firstly, the event number of each pixel of the HR DVS image is generated by obtaining the HR event-count map (ECM) from the LR DVS recording with a sparse representation based method. The rate function over time line of the point process of each HR pixel is computed using a spatiotemporal filter on the corresponding LR neighbor pixels. Secondly, the event sequence of each new pixel is obtained with a thinning based event sampling algorithm. A metric is proposed to assess the event-stream SR quality. The effectiveness of the proposed method is demonstrated through obtaining HR event-stream images from a series of DVS recordings. This work enables many potential.

Vector function optimization problems, in which one or more variables are multidimensional vectors or infinite-dimensional vectors, have been extensively studied and demonstrated in existing schemes. In various real life applications, a cost function to be optimized usually involves matrix variables subjected to certain constraints. Locating its minimum can be modeled as an optimization problem on matrix manifold which is investigated in this paper. We first present an index-notation-arrangement based chain rule (I-Chain rule) to obtain the gradient information of the cost function. Two iterative algorithms, namely, trace-constraint-based projected gradient method (TPGM) and orthonormal-constraint-based projected gradient method (OPGM) are proposed and their convergence properties are established. We find that the network control problems can be effectively solved by both TPGM and OPGM. Two important phenomena are observed. For controlling directed networks with selectable inputs, both TPGM and OPGM tend to locate the nodes that divides the elementary stem/circle/dilation equally for consuming less energy, with OPGM having a slightly higher chance than TPGM. For controlling directed networks by only evolving the connection strengths on a fixed network structure, we find that after a network adaptively changes its topology in such a way that many similar sub-networks are gradually evolved, the control cost attains its minimum. Our work takes a further step from understanding optimization problems on matrix manifold to extending their applications in science and engineering.

Different from traditional frame-based cameras, event-based dynamic vision sensor (DVS) converts the visual information into spatiotemporal event streams. Convolutional neural networks (CNNs) have recently achieved outstanding classification performance while require a very large number of annotated samples. However, a lack of available large-scale event-stream datasets prevents application of CNNs to classification of such event streams. In this work, we show how the deep representation learned with an originally optimized CNN is efficiently transferred to the event-stream classification tasks. In our classification method, a spike-event temporal coding is used to encoding the spike-event information of each pixel. This temporal coding mechanism is implemented based on the subthreshold dynamic of the leaky integrate-and-fire (LIF) model. Three popular event-stream datasets were used to evaluate the performance of the proposed method. Results show that the proposed method leads to significantly improved classification accuracy, outperforming the current state of the art methods on the three event-stream datasets. Besides, the robustness of our method was verified in the MNIST-DVS dataset when Gaussian temporal noises were added to the timestamps of the events. Finally, we find that fine tuning with a small amount of event-stream data would improve the classification performance. This work can be easily extended to more complex scenarios and more fascinating and potential visual applications.

Deep neural networks have been achieving booming breakthroughs in various artificial intelligence tasks, however they are notorious for consuming unbearable hardware resources, training time and power. The emerging pruning/binarization methods, which aim at both decreasing overheads and retaining high performance, seem to promise applications on portable devices. However, even with these most advanced algorithms, we have to save the full-precision weights during the gradient descent process which remains size and power bottlenecks of memory access and the resulting computation. To address this challenge, we propose a unified discrete state transition (DST) framework by introducing a probabilistic projection operator that constrains the weight matrices in a discrete weight space (DWS) with configurable number of states, throughout the whole training process. The experimental results over various data sets including MNIST, CIFAR10 and SVHN show the effectiveness of this framework. The direct transition between discrete states significantly saves memory for storing weights in full precision, as well as simplifies the computation of weight updating. The proposed DST framework is hardware friendly as it can be easily implemented by a wide range of emerging portable devices, including binary, ternary and multiple-level memory devices. This work paves the way for on-chip learning on various portable devices in the near future.

Persistent activity holds the transient stimulus for up to many seconds even after the stimulus is gone. It has been implemented in a class of models known as continuous attractor neural networks, which have infinite stable states corresponding to persistent activity patterns. Continuous attractor neural network remains stable so does not change systematically in the absence of stimulus input. Continuous attractor is a set of connected stable equilibrium points and has been used to describe the storing of continuous stimuli in neural networks. The background input of the networks plays an important role in continuous attractor neural network. In this paper, dynamical properties of continuous attractor neural network with two background input tuning schemes are investigated: constant input shifting and oscillation background activity. Simulations are employed to illustrate the theory.

Thermal stability of 100 nm Ge 2 Sb 2 Te 5 thin film during annealing from room temperature to 240 °C inside a UHV chamber was studied in situ by X-ray photoelectron spectroscopy (XPS) and ex situ by X-ray diffraction (XRD) and atomic force microscopy (AFM). Ge species are found to diffuse preferentially to the surface when GST film is annealed from 25 °C to 100 °C. This process is accompanied by a change of phase whereby the amorphous film completely becomes face-center-cubic (FCC) phase at 100 °C. From 100 °C to 200 °C, both Sb and Te species are observed to diffuse more to the surface. The FCC phase is partially changed into hexagonal-close-pack (HCP) phase at 200 °C. At 220 °C, FCC phase is completely transformed into HCP phase. Loss of Sb and Te are also detected from the surface and this is attributed to desorption due to their high vapor pressures. At 240 °C, Sb and Te species are found to have desorbed completely from the surface, and leave behind Ge-rich 3D droplets on the surface. The separation of Ge 2 Sb 2 Te 5 into Sb,Te-rich phase and Ge-rich phase is thus the main mechanism to account for the failure of Ge 2 Sb 2 Te 5 -based phase change memory devices under thermal stress.

Cleaning the surfaces of the as-deposited Ge 2 Sb 2 Te 5 was studied by X-ray photoelectron spectroscopy (XPS), atomic force microscopy (AFM) and X-ray diffraction (XRD). The mixed native oxides on the as-deposited Ge 2 Sb 2 Te 5 surface can be easily removed by dipping Ge 2 Sb 2 Te 5 in de-ionized water for 1 min, while the surface morphology remains unchanged after cleaning. Native oxides only re-grow after exposure to air for more than 4 min. Although dipping in water leads to a surface layer deficient in Ge and Sb, the surface composition of Ge 2 Sb 2 Te 5 can recover to its stoichiometric value after annealing at 200 °C in vacuum. The phase remains amorphous at room temperature after dipping in water, and changes to fcc and hcp after annealing at 100 and 220 °C, respectively.

The speed of sound of the gaseous trifluoroiodomethane (CF3I) was measured for temperatures from 273.15 to 333.15 K and pressures from 58 to 276 kPa with a cylindrical, variable-path acoustic interferometer operating at 156.252 kHz. The uncertainty of the speed of sound was less than ±0.1%. The ideal-gas heat capacity at constant pressure and the second acoustic virial coefficients were determined over the temperature range from the speed of sound measurements. The ideal-gas heat capacity at constant pressure and second virial coefficients calculated from the present speed of sound measurements were compared with the available data. The uncertainty of the ideal-gas heat capacity at constant pressure was estimated to be less than ± 1%.