On the capabilities and computational costs of neuron models

Toward the Optimal Design and FPGA Implementation of Spiking Neural Networks

The performance of a biologically plausible spiking neural network (SNN) largely depends on the model parameters and neural dynamics. This article proposes a parameter optimization scheme for improving the performance of a biologically plausible SNN and a parallel on-field-programmable gate array (FPGA) online learning neuromorphic platform for the digital implementation based on two numerical methods, namely, the Euler and third-order Runge-Kutta (RK3) methods. The optimization scheme explores the impact of biological time constants on information transmission in the SNN and improves the convergence rate of the SNN on digit recognition with a suitable choice of the time constants. The parallel digital implementation leads to a significant speedup over software simulation on a general-purpose CPU. The parallel implementation with the Euler method enables around 180× ( 20× ) training (inference) speedup over a Pytorch-based SNN simulation on CPU. Moreover, compared with previous work, our parallel implementation shows more than 300× ( 240× ) improvement on speed and 180× ( 250× ) reduction in energy consumption for training (inference). In addition, due to the high-order accuracy, the RK3 method is demonstrated to gain 2× training speedup over the Euler method, which makes it suitable for online training in real-time applications.

Memristive Izhikevich Spiking Neuron Model and Its Application in Oscillatory Associative Memory

The Izhikevich (IZH) spiking neuron model can display spiking and bursting behaviors of neurons. Based on the switching property and bio-plausibility of the memristor, the memristive Izhikevich (MIZH) spiking neuron model is built. Firstly, the MIZH spiking model is introduced and used to generate 23 spiking patterns. We compare the 23 spiking patterns produced by the IZH and MIZH spiking models. Secondly, the MIZH spiking model actively reproduces various neuronal behaviors, including the excitatory cortical neurons, the inhibitory cortical neurons, and other cortical neurons. Finally, the collective dynamic activities of the MIZH neuronal network are performed, and the MIZH oscillatory network is constructed. Experimental results illustrate that the constructed MIZH spiking neuron model performs high firing frequency and good frequency adaptation. The model can easily simulate various spiking and bursting patterns of distinct neurons in the brain. The MIZH neuronal network realizes the synchronous and asynchronous collective behaviors. The MIZH oscillatory network can memorize and retrieve the information patterns correctly and efficiently with high retrieval accuracy.

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.

Active Hearing Mechanisms Inspire Adaptive Amplification in an Acoustic Sensor System

Over many millions of years of evolution, nature has developed some of the most adaptable sensors and sensory systems possible, capable of sensing, conditioning and processing signals in a very power- and size-effective manner. By looking into biological sensors and systems as a source of inspiration, this paper presents the study of a bioinspired concept of signal processing at the sensor level. By exploiting a feedback control mechanism between a front-end acoustic receiver and back-end neuronal based computation, a nonlinear amplification with hysteretic behavior is created. Moreover, the transient response of the front-end acoustic receiver can also be controlled and enhanced. A theoretical model is proposed and the concept is prototyped experimentally through an embedded system setup that can provide dynamic adaptations of a sensory system comprising a MEMS microphone placed in a closed-loop feedback system. It faithfully mimics the mosquito's active hearing response as a function of the input sound intensity. This is an adaptive acoustic sensor system concept that can be exploited by sensor and system designers within acoustics and ultrasonic engineering fields.

Physiological models of the lateral superior olive

In computational biology, modeling is a fundamental tool for formulating, analyzing and predicting complex phenomena. Most neuron models, however, are designed to reproduce certain small sets of empirical data. Hence their outcome is usually not compatible or comparable with other models or datasets, making it unclear how widely applicable such models are. In this study, we investigate these aspects of modeling, namely credibility and generalizability, with a specific focus on auditory neurons involved in the localization of sound sources. The primary cues for binaural sound localization are comprised of interaural time and level differences (ITD/ILD), which are the timing and intensity differences of the sound waves arriving at the two ears. The lateral superior olive (LSO) in the auditory brainstem is one of the locations where such acoustic information is first computed. An LSO neuron receives temporally structured excitatory and inhibitory synaptic inputs that are driven by ipsi- and contralateral sound stimuli, respectively, and changes its spike rate according to binaural acoustic differences. Here we examine seven contemporary models of LSO neurons with different levels of biophysical complexity, from predominantly functional ones (‘shot-noise’ models) to those with more detailed physiological components (variations of integrate-and-fire and Hodgkin-Huxley-type). These models, calibrated to reproduce known monaural and binaural characteristics of LSO, generate largely similar results to each other in simulating ITD and ILD coding. Our comparisons of physiological detail, computational efficiency, predictive performances, and further expandability of the models demonstrate (1) that the simplistic, functional LSO models are suitable for applications where low computational costs and mathematical transparency are needed, (2) that more complex models with detailed membrane potential dynamics are necessary for simulation studies where sub-neuronal nonlinear processes play important roles, and (3) that, for general purposes, intermediate models might be a reasonable compromise between simplicity and biological plausibility.

Computational models help our understanding of complex biological systems, by identifying their key elements and revealing their operational principles. Close comparisons between model predictions and empirical observations ensure our confidence in a model as a building block for further applications. Most current neuronal models, however, are constructed to replicate only a small specific set of experimental data. Thus, it is usually unclear how these models can be generalized to different datasets and how they compare with each other. In this paper, seven neuronal models are examined that are designed to reproduce known physiological characteristics of auditory neurons involved in the detection of sound source location. Despite their different levels of complexity, the models generate largely similar results when their parameters are tuned with common criteria. Comparisons show that simple models are computationally more efficient and theoretically transparent, and therefore suitable for rigorous mathematical analyses and engineering applications including real-time simulations. In contrast, complex models are necessary for investigating the relationship between underlying biophysical processes and sub- and suprathreshold spiking properties, although they have a large number of unconstrained, unverified parameters. Having identified their advantages and drawbacks, these auditory neuron models may readily be used for future studies and applications.

Computational modeling of peripheral pain: a commentary

This commentary is intended to find possible explanations for the low impact of computational modeling on pain research. We discuss the main strategies that have been used in building computational models for the study of pain. The analysis suggests that traditional models lack biological plausibility at some levels, they do not provide clinically relevant results, and they cannot capture the stochastic character of neural dynamics. On this basis, we provide some suggestions that may be useful in building computational models of pain with a wider range of applications.