Accuracy evaluation of numerical methods used in state-of-the-art simulators for spiking neural networks

Supervised learning in spiking neural networks: A review of algorithms and evaluations

As a new brain-inspired computational model of the artificial neural network, a spiking neural network encodes and processes neural information through precisely timed spike trains. Spiking neural networks are composed of biologically plausible spiking neurons, which have become suitable tools for processing complex temporal or spatiotemporal information. However, because of their intricately discontinuous and implicit nonlinear mechanisms, the formulation of efficient supervised learning algorithms for spiking neural networks is difficult, and has become an important problem in this research field. This article presents a comprehensive review of supervised learning algorithms for spiking neural networks and evaluates them qualitatively and quantitatively. First, a comparison between spiking neural networks and traditional artificial neural networks is provided. The general framework and some related theories of supervised learning for spiking neural networks are then introduced. Furthermore, the state-of-the-art supervised learning algorithms in recent years are reviewed from the perspectives of applicability to spiking neural network architecture and the inherent mechanisms of supervised learning algorithms. A performance comparison of spike train learning of some representative algorithms is also made. In addition, we provide five qualitative performance evaluation criteria for supervised learning algorithms for spiking neural networks and further present a new taxonomy for supervised learning algorithms depending on these five performance evaluation criteria. Finally, some future research directions in this research field are outlined.

On the accuracy and computational cost of spiking neuron implementation

Since more than a decade ago, three statements about spiking neuron (SN) implementations have been widely accepted: 1) Hodgkin and Huxley (HH) model is computationally prohibitive, 2) Izhikevich (IZH) artificial neuron is as efficient as Leaky Integrate-and-Fire (LIF) model, and 3) IZH model is more efficient than HH model (Izhikevich, 2004). As suggested by Hodgkin and Huxley (1952), their model operates in two modes: by using the α's and β's rate functions directly (HH model) and by storing them into tables (HHT model) for computational cost reduction. Recently, it has been stated that: 1) HHT model (HH using tables) is not prohibitive, 2) IZH model is not efficient, and 3) both HHT and IZH models are comparable in computational cost (Skocik & Long, 2014). That controversy shows that there is no consensus concerning SN simulation capacities. Hence, in this work, we introduce a refined approach, based on the multiobjective optimization theory, describing the SN simulation capacities and ultimately choosing optimal simulation parameters. We have used normalized metrics to define the capacity levels of accuracy, computational cost, and efficiency. Normalized metrics allowed comparisons between SNs at the same level or scale. We conducted tests for balanced, lower, and upper boundary conditions under a regular spiking mode with constant and random current stimuli. We found optimal simulation parameters leading to a balance between computational cost and accuracy. Importantly, and, in general, we found that 1) HH model (without using tables) is the most accurate, computationally inexpensive, and efficient, 2) IZH model is the most expensive and inefficient, 3) both LIF and HHT models are the most inaccurate, 4) HHT model is more expensive and inaccurate than HH model due to α's and β's table discretization, and 5) HHT model is not comparable in computational cost to IZH model. These results refute the theory formulated over a decade ago (Izhikevich, 2004) and go more in-depth in the statements formulated by Skocik and Long (2014). Our statements imply that the number of dimensions or FLOPS in the SNs are theoretical but not practical indicators of the true computational cost. The metric we propose for the computational cost is more precise than FLOPS and was found to be invariant to computer architecture. Moreover, we found that the firing frequency used in previous works is a necessary but an insufficient metric to evaluate the simulation accuracy. We also show that our results are consistent with the theory of numerical methods and the theory of SN discontinuity. Discontinuous SNs, such LIF and IZH models, introduce a considerable error every time a spike is generated. In addition, compared to the constant input current, the random input current increases the computational cost and inaccuracy. Besides, we found that the search for optimal simulation parameters is problem-specific. That is important because most of the previous works have intended to find a general and unique optimal simulation. Here, we show that this solution could not exist because it is a multiobjective optimization problem that depends on several factors. This work sets up a renewed thesis concerning the SN simulation that is useful to several related research areas, including the emergent Deep Spiking Neural Networks.

Performance Comparison of the Digital Neuromorphic Hardware SpiNNaker and the Neural Network Simulation Software NEST for a Full-Scale Cortical Microcircuit Model

The digital neuromorphic hardware SpiNNaker has been developed with the aim of enabling large-scale neural network simulations in real time and with low power consumption. Real-time performance is achieved with 1 ms integration time steps, and thus applies to neural networks for which faster time scales of the dynamics can be neglected. By slowing down the simulation, shorter integration time steps and hence faster time scales, which are often biologically relevant, can be incorporated. We here describe the first full-scale simulations of a cortical microcircuit with biological time scales on SpiNNaker. Since about half the synapses onto the neurons arise within the microcircuit, larger cortical circuits have only moderately more synapses per neuron. Therefore, the full-scale microcircuit paves the way for simulating cortical circuits of arbitrary size. With approximately 80, 000 neurons and 0.3 billion synapses, this model is the largest simulated on SpiNNaker to date. The scale-up is enabled by recent developments in the SpiNNaker software stack that allow simulations to be spread across multiple boards. Comparison with simulations using the NEST software on a high-performance cluster shows that both simulators can reach a similar accuracy, despite the fixed-point arithmetic of SpiNNaker, demonstrating the usability of SpiNNaker for computational neuroscience applications with biological time scales and large network size. The runtime and power consumption are also assessed for both simulators on the example of the cortical microcircuit model. To obtain an accuracy similar to that of NEST with 0.1 ms time steps, SpiNNaker requires a slowdown factor of around 20 compared to real time. The runtime for NEST saturates around 3 times real time using hybrid parallelization with MPI and multi-threading. However, achieving this runtime comes at the cost of increased power and energy consumption. The lowest total energy consumption for NEST is reached at around 144 parallel threads and 4.6 times slowdown. At this setting, NEST and SpiNNaker have a comparable energy consumption per synaptic event. Our results widen the application domain of SpiNNaker and help guide its development, showing that further optimizations such as synapse-centric network representation are necessary to enable real-time simulation of large biological neural networks.

A unified framework for spiking and gap-junction interactions in distributed neuronal network simulations

Contemporary simulators for networks of point and few-compartment model neurons come with a plethora of ready-to-use neuron and synapse models and support complex network topologies. Recent technological advancements have broadened the spectrum of application further to the efficient simulation of brain-scale networks on supercomputers. In distributed network simulations the amount of spike data that accrues per millisecond and process is typically low, such that a common optimization strategy is to communicate spikes at relatively long intervals, where the upper limit is given by the shortest synaptic transmission delay in the network. This approach is well-suited for simulations that employ only chemical synapses but it has so far impeded the incorporation of gap-junction models, which require instantaneous neuronal interactions. Here, we present a numerical algorithm based on a waveform-relaxation technique which allows for network simulations with gap junctions in a way that is compatible with the delayed communication strategy. Using a reference implementation in the NEST simulator, we demonstrate that the algorithm and the required data structures can be smoothly integrated with existing code such that they complement the infrastructure for spiking connections. To show that the unified framework for gap-junction and spiking interactions achieves high performance and delivers high accuracy in the presence of gap junctions, we present benchmarks for workstations, clusters, and supercomputers. Finally, we discuss limitations of the novel technology.

Reducing the computational footprint for real-time BCPNN learning

The implementation of synaptic plasticity in neural simulation or neuromorphic hardware is usually very resource-intensive, often requiring a compromise between efficiency and flexibility. A versatile, but computationally-expensive plasticity mechanism is provided by the Bayesian Confidence Propagation Neural Network (BCPNN) paradigm. Building upon Bayesian statistics, and having clear links to biological plasticity processes, the BCPNN learning rule has been applied in many fields, ranging from data classification, associative memory, reward-based learning, probabilistic inference to cortical attractor memory networks. In the spike-based version of this learning rule the pre-, postsynaptic and coincident activity is traced in three low-pass-filtering stages, requiring a total of eight state variables, whose dynamics are typically simulated with the fixed step size Euler method. We derive analytic solutions allowing an efficient event-driven implementation of this learning rule. Further speedup is achieved by first rewriting the model which reduces the number of basic arithmetic operations per update to one half, and second by using look-up tables for the frequently calculated exponential decay. Ultimately, in a typical use case, the simulation using our approach is more than one order of magnitude faster than with the fixed step size Euler method. Aiming for a small memory footprint per BCPNN synapse, we also evaluate the use of fixed-point numbers for the state variables, and assess the number of bits required to achieve same or better accuracy than with the conventional explicit Euler method. All of this will allow a real-time simulation of a reduced cortex model based on BCPNN in high performance computing. More important, with the analytic solution at hand and due to the reduced memory bandwidth, the learning rule can be efficiently implemented in dedicated or existing digital neuromorphic hardware.

Limits to high-speed simulations of spiking neural networks using general-purpose computers

To understand how the central nervous system performs computations using recurrent neuronal circuitry, simulations have become an indispensable tool for theoretical neuroscience. To study neuronal circuits and their ability to self-organize, increasing attention has been directed toward synaptic plasticity. In particular spike-timing-dependent plasticity (STDP) creates specific demands for simulations of spiking neural networks. On the one hand a high temporal resolution is required to capture the millisecond timescale of typical STDP windows. On the other hand network simulations have to evolve over hours up to days, to capture the timescale of long-term plasticity. To do this efficiently, fast simulation speed is the crucial ingredient rather than large neuron numbers. Using different medium-sized network models consisting of several thousands of neurons and off-the-shelf hardware, we compare the simulation speed of the simulators: Brian, NEST and Neuron as well as our own simulator Auryn. Our results show that real-time simulations of different plastic network models are possible in parallel simulations in which numerical precision is not a primary concern. Even so, the speed-up margin of parallelism is limited and boosting simulation speeds beyond one tenth of real-time is difficult. By profiling simulation code we show that the run times of typical plastic network simulations encounter a hard boundary. This limit is partly due to latencies in the inter-process communications and thus cannot be overcome by increased parallelism. Overall, these results show that to study plasticity in medium-sized spiking neural networks, adequate simulation tools are readily available which run efficiently on small clusters. However, to run simulations substantially faster than real-time, special hardware is a prerequisite.

Toward unified hybrid simulation techniques for spiking neural networks

In the field of neural network simulation techniques, the common conception is that spiking neural network simulators can be divided in two categories: time-step-based and event-driven methods. In this letter, we look at state-of-the art simulation techniques in both categories and show that a clear distinction between both methods is increasingly difficult to define. In an attempt to improve the weak points of each simulation method, ideas of the alternative method are, sometimes unknowingly, incorporated in the simulation engine. Clearly the ideal simulation method is a mix of both methods. We formulate the key properties of such an efficient and generally applicable hybrid approach.

Creating, documenting and sharing network models

As computational neuroscience matures, many simulation environments are available that are useful for neuronal network modeling. However, methods for successfully documenting models for publication and for exchanging models and model components among these projects are still under development. Here we briefly review existing software and applications for network model creation, documentation and exchange. Then we discuss a few of the larger issues facing the field of computational neuroscience regarding network modeling and suggest solutions to some of these problems, concentrating in particular on standardized network model terminology, notation, and descriptions and explicit documentation of model scaling. We hope this will enable and encourage computational neuroscientists to share their models more systematically in the future.