All issues

Working memory mechanisms in spiking neural networks consisting of leaky integrate-and-fire neurons with adaptive threshold and synaptic plasticity are studied in this work. Moderate size networks including thousands of neurons were explored. Working memory is a network ability to keep in its state the information about recent stimuli presented to the network such that this information is sufficient to determine which stimulus has been presented. In this study, network state is defined as the current characteristics of network activity only — without internal state of its neurons. In order to discover the neuronal structures serving as a possible substrate of the memory mechanism, optimization of the network parameters and structure using genetic algorithm was carried out. Two kinds of neuronal structures with the desired properties were found. These are neuron pairs mutually connected by strong synaptic links and long tree-like neuronal ensembles. It was shown that only the neuron pairs are suitable for efficient and reliable implementation of working memory. Properties of such memory units and structures formed by them are explored in the present study. It is shown that characteristics of the studied two-neuron memory units can be set easily by the respective choice of the parameters of its neurons and synaptic connections. Besides that, this work demonstrates that ensembles of these structures can provide the network with capability of unsupervised learning to recognize patterns in the input signal.

The grid-characteristic method is successfully used for solving hyperbolic systems of partial differential equations (for example, transport / acoustic / elastic equations). It allows to construct correctly algorithms on contact boundaries and boundaries of the integration domain, to a certain extent to take into account the physics of the problem (propagation of discontinuities along characteristic curves), and has the property of monotonicity, which is important for considered problems. In the cases of two-dimensional and three-dimensional problems the method makes use of a coordinate splitting technique, which enables us to solve the original equations by solving several one-dimensional ones consecutively. It is common to use up to 3-rd order one-dimensional schemes with simple splitting techniques which do not allow for the convergence order to be higher than two (with respect to time). Significant achievements in the operator splitting theory were done, the existence of higher-order schemes was proved. Its peculiarity is the need to perform a step in the opposite direction in time, which gives rise to difficulties, for example, for parabolic problems.

In this work coordinate splitting of the 3-rd and 4-th order were used for the two-dimensional hyperbolic problem of the linear elasticity. This made it possible to increase the final convergence order of the computational algorithm. The paper empirically estimates the convergence in L1 and L∞ norms using analytical solutions of the system with the sufficient degree of smoothness. To obtain objective results, we considered the cases of longitudinal and transverse plane waves propagating both along the diagonal of the computational cell and not along it. Numerical experiments demonstrated the improved accuracy and convergence order of constructed schemes. These improvements are achieved with the cost of three- or fourfold increase of the computational time (for the 3-rd and 4-th order respectively) and no additional memory requirements. The proposed improvement of the computational algorithm preserves the simplicity of its parallel implementation based on the spatial decomposition of the computational grid.

The method for mapping of intermediate representations (IR) set of C, C++ programs to vector embedding space is considered to create an empirical estimation framework for static performance prediction using LLVM compiler infrastructure. The usage of embeddings makes programs easier to compare due to avoiding Control Flow Graphs (CFG) and Data Flow Graphs (DFG) direct comparison. This method is based on transformation series of the initial IR such as: instrumentation — injection of artificial instructions in an instrumentation compiler’s pass depending on load offset delta in the current instruction compared to the previous one, mapping of instrumented IR into multidimensional vector with IR2Vec and dimension reduction with t-SNE (t-distributed stochastic neighbor embedding) method. The D1 cache miss ratio measured with perf stat tool is considered as performance metric. A heuristic criterion of programs having more or less cache miss ratio is given. This criterion is based on embeddings of programs in 2D-space. The instrumentation compiler’s pass developed in this work is described: how it generates and injects artificial instructions into IR within the used memory model. The software pipeline that implements the performance estimation based on LLVM compiler infrastructure is given. Computational experiments are performed on synthetic tests which are the sets of programs with the same CFGs but with different sequences of offsets used when accessing the one-dimensional array of a given size. The correlation coefficient between performance metric and distance to the worst program’s embedding is measured and proved to be negative regardless of t-SNE initialization. This fact proves the heuristic criterion to be true. The process of such synthetic tests generation is also considered. Moreover, the variety of performance metric in programs set in such a test is proposed as a metric to be improved with exploration of more tests generators.

Modeling of carbon nanostructures by means of classical molecular dynamics requires a lot of computations. One of the ways to improve the performance of basic algorithms is to transform them for running on SIMD-type computing systems such as systems with dedicated GPU. In this work we describe the development of algorithms for computation of many-body interaction based on Tersoff and embedded-atom potentials by means of OpenCL technology. OpenCL standard provides universality and portability of the algorithms and can be successfully used for development of the software for heterogeneous computing systems. The performance of algorithms is evaluated on CPU and GPU hardware platforms. It is shown that concurrent memory writes is effective for Tersoff bond order potential. The same approach for embedded-atom potential is shown to be slower than algorithm without concurrent memory access. Performance evaluation shows a significant GPU acceleration of energy-force evaluation algorithms for many-body potentials in comparison to the corresponding serial implementations.