Hardware implementation of stochastic spiking neural networks

Spiking Neural Networks, the last generation of Artificial Neural Networks, are characterized by its bio-inspired nature and by a higher computational capacity with respect to other neural models. In real biological neurons, stochastic processes represent an important mechanism of neural behavior and are responsible of its special arithmetic capabilities. In this work we present a simple hardware implementation of spiking neurons that considers this probabilistic nature. The advantage of the proposed implementation is that it is fully digital and therefore can be massively implemented in Field Programmable Gate Arrays. The high computational capabilities of the proposed model are demonstrated by the study of both feed-forward and recurrent networks that are able to implement high-speed signal filtering and to solve complex systems of linear equations.     

VLSI implementation of neural networks

Currently, fuzzy controllers are the most popular choice for hardware implementation of complex control surfaces because they are easy to design. Neural controllers are more complex and hard to train, but provide an outstanding control surface with much less error than that of a fuzzy controller. There are also some problems that have to be solved before the networks can be implemented on VLSI chips. First, an approximation function needs to be developed because CMOS neural networks have an activation function different than any function used in neural network software. Next, this function has to be used to train the network. Finally, the last problem for VLSI designers is the quantization effect caused by discrete values of the channel length (L) and width (W) of MOS transistor geometries. Two neural networks were designed in 1.5 microm technology. Using adequate approximation functions solved the problem of activation function. With this approach, trained networks were characterized by very small errors. Unfortunately, when the weights were quantized, errors were increased by an order of magnitude. However, even though the errors were enlarged, the results obtained from neural network hardware implementations were superior to the results obtained with fuzzy system approach.     

Field programmable chemistry: integrated chemical and electronic processing of informational molecules towards electronic chemical cells

The topic addressed is that of combining self-constructing chemical systems with electronic computation to form unconventional embedded computation systems performing complex nano-scale chemical tasks autonomously. The hybrid route to complex programmable chemistry, and ultimately to artificial cells based on novel chemistry, requires a solution of the two-way massively parallel coupling problem between digital electronics and chemical systems. We present a chemical microprocessor technology and show how it can provide a generic programmable platform for complex molecular processing tasks in Field Programmable Chemistry, including steps towards the grand challenge of constructing the first electronic chemical cells. Field programmable chemistry employs a massively parallel field of electrodes, under the control of latched voltages, which are used to modulate chemical activity. We implement such a field programmable chemistry which links to chemistry in rather generic, two-phase microfluidic channel networks that are separated into weakly coupled domains. Electric fields, produced by the high-density array of electrodes embedded in the channel floors, are used to control the transport of chemicals across the hydrodynamic barriers separating domains. In the absence of electric fields, separate microfluidic domains are essentially independent with only slow diffusional interchange of chemicals. Electronic chemical cells, based on chemical microprocessors, exploit a spatially resolved sandwich structure in which the electronic and chemical systems are locally coupled through homogeneous fine-grained actuation and sensor networks and play symmetric and complementary roles. We describe how these systems are fabricated, experimentally test their basic functionality, simulate their potential (e.g. for feed forward digital electrophoretic (FFDE) separation) and outline the application to building electronic chemical cells.     

Advancing simulations of biological materials: applications of coarse-grained models on graphics processing unit hardware

The timescales of biological processes, primarily those inherent to the molecular mechanisms of disease, are long (>μs) and involve complex interactions of systems consisting of many atoms (>10⁶). Simulating these systems requires an advanced computational approach, and as such, coarse-grained (CG) models have been developed and highly optimised for accelerator hardware, primarily graphics processing units (GPUs). In this review, I discuss the implementation of CG models for biologically relevant systems, and show how such models can be optimised and perform well on GPU-accelerated hardware. Several examples of GPU implementations of CG models for both molecular dynamics and Monte Carlo simulations on purely GPU and hybrid CPU/GPU architectures are presented. Both the hardware and algorithmic limitations of various models, which depend greatly on the application of interest, are discussed.     

Automating parallel implementation of neural learning algorithms

Neural learning algorithms generally involve a number of identical processing units, which are fully or partially connected, and involve an update function, such as a ramp, a sigmoid or a Gaussian function for instance. Some variations also exist, where units can be heterogeneous, or where an alternative update technique is employed, such as a pulse stream generator. Associated with connections are numerical values that must be adjusted using a learning rule, and and dictated by parameters that are learning rule specific, such as momentum, a learning rate, a temperature, amongst others. Usually, neural learning algorithms involve local updates, and a global interaction between units is often discouraged, except in instances where units are fully connected, or involve synchronous updates. In all of these instances, concurrency within a neural algorithm cannot be fully exploited without a suitable implementation strategy. A design scheme is described for translating a neural learning algorithm from inception to implementation on a parallel machine using PVM or MPI libraries, or onto programmable logic such as FPGAs. A designer must first describe the algorithm using a specialised Neural Language, from which a Petri net (PN) model is constructed automatically for verification, and building a performance model. The PN model can be used to study issues such as synchronisation points, resource sharing and concurrency within a learning rule. Specialised constructs are provided to enable a designer to express various aspects of a learning rule, such as the number and connectivity of neural nodes, the interconnection strategies, and information flows required by the learning algorithm. A scheduling and mapping strategy is then used to translate this PN model onto a multiprocessor template. We demonstrate our technique using a Kohonen and backpropagation learning rules, implemented on a loosely coupled workstation cluster, and a dedicated parallel machine, with PVM libraries.     

Computer vision system for superpixel classification and segmentation of sheep

This paper presents an experiment with four different convolutional neural networks architectures that aim to classify segments of a sheep using a dataset of superpixels. The proposal used an image dataset with 512 images of 32 sheep. In this dataset of images, we applied the Simple Linear Iterative Clustering technique with a K number parameter of 600 to generate the dataset of superpixels that was later processed in the deep neural networks. We selected four architectures for training the models: VGG16, ResNet152V2, InceptionV3, and DenseNet201. The experiment was conducted using cross-validation with five-folds separating the dataset into training (60%), validation (20%), and testing (20%). The best result presented was from the approach with the DenseNet201 technique with an F-score of 0.928. When applying ANOVA, the P-value was 0.0000000000329 (3.29e-11***) between the tested architectures, which shows that the results are statistically significant. Therefore, DenseNet201 presented itself as a relevant architecture for this problem that aims to classify the superpixels referring to a sheep and the image's background, and the average IoU with post-processing for image segmentation with DenseNet201 obtained 0.8332. Thus, we can highlight the contributions of this research as a methodology to segment images of mixed-breed sheep of the Texel and Santa Inês breeds using convolutional neural networks and provide a non-invasive method that can support other implementations such as animal tracking and weight prediction.     

Neural networks for perceptual processing: from simulation tools to theories

Neural networks are modelling tools that are, in principle, able to capture the input-output behaviour of arbitrary systems that may include the dynamics of animal populations or brain circuits. While a neural network model is useful if it captures phenomenologically the behaviour of the target system in this way, its utility is amplified if key mechanisms of the model can be discovered, and identified with those of the underlying system. In this review, we first describe, at a fairly high level with minimal mathematics, some of the tools used in constructing neural network models. We then go on to discuss the implications of network models for our understanding of the system they are supposed to describe, paying special attention to those models that deal with neural circuits and brain systems. We propose that neural nets are useful for brain modelling if they are viewed in a wider computational framework originally devised by Marr. Here, neural networks are viewed as an intermediate mechanistic abstraction between 'algorithm' and 'implementation', which can provide insights into biological neural representations and their putative supporting architectures.     

Towards a more general understanding of the algorithmic utility of recurrent connections

Lateral and recurrent connections are ubiquitous in biological neural circuits. Yet while the strong computational abilities of feedforward networks have been extensively studied, our understanding of the role and advantages of recurrent computations that might explain their prevalence remains an important open challenge. Foundational studies by Minsky and Roelfsema argued that computations that require propagation of global information for local computation to take place would particularly benefit from the sequential, parallel nature of processing in recurrent networks. Such “tag propagation” algorithms perform repeated, local propagation of information and were originally introduced in the context of detecting connectedness, a task that is challenging for feedforward networks. Here, we advance the understanding of the utility of lateral and recurrent computation by first performing a large-scale empirical study of neural architectures for the computation of connectedness to explore feedforward solutions more fully and establish robustly the importance of recurrent architectures. In addition, we highlight a tradeoff between computation time and performance and construct hybrid feedforward/recurrent models that perform well even in the presence of varying computational time limitations. We then generalize tag propagation architectures to propagating multiple interacting tags and demonstrate that these are efficient computational substrates for more general computations of connectedness by introducing and solving an abstracted biologically inspired decision-making task. Our work thus clarifies and expands the set of computational tasks that can be solved efficiently by recurrent computation, yielding hypotheses for structure in population activity that may be present in such tasks.     

Neural networks in analog hardware--design and implementation issues

This paper presents a brief review of some analog hardware implementations of neural networks. Several criteria for the classification of general neural networks implementations are discussed and a taxonomy induced by these criteria is presented. The paper also discusses some characteristics of analog implementations as well as some trade-offs and issues identified in the work reviewed. Parameters such as precision, chip area, power consumption, speed and noise susceptibility are discussed in the context of neural implementations. A unified review of various "VLSI friendly" algorithms is also presented. The paper concludes with some conclusions drawn from the analysis of the implementations presented.     

Internal representations for associative memory

We describe a class of feed forward neural network models for associative content addressable memory (ACAM) which utilize sparse internal representations for stored data. In addition to the input and output layers, our networks incorporate an intermediate processing layer which serves to label each stored memory and to perform error correction and association. We study two classes of internal label representations: the unary representation and various sparse, distributed representations. Finally, we consider storage of sparse data and sparsification of data. These models are found to have advantages in terms of storage capacity, hardware efficiency, and recall reliability when compared to the Hopfield model, and to possess analogies to both biological neural networks and standard digital computer memories.     

A transition-constrained discrete hidden Markov model for automatic sleep staging

Background

Principal component analysis (PCA) has been widely employed for automatic neuronal spike sorting. Calculating principal components (PCs) is computationally expensive, and requires complex numerical operations and large memory resources. Substantial hardware resources are therefore needed for hardware implementations of PCA. General Hebbian algorithm (GHA) has been proposed for calculating PCs of neuronal spikes in our previous work, which eliminates the needs of computationally expensive covariance analysis and eigenvalue decomposition in conventional PCA algorithms. However, large memory resources are still inherently required for storing a large volume of aligned spikes for training PCs. The large size memory will consume large hardware resources and contribute significant power dissipation, which make GHA difficult to be implemented in portable or implantable multi-channel recording micro-systems.

Method

In this paper, we present a new algorithm for PCA-based spike sorting based on GHA, namely stream-based Hebbian eigenfilter, which eliminates the inherent memory requirements of GHA while keeping the accuracy of spike sorting by utilizing the pseudo-stationarity of neuronal spikes. Because of the reduction of large hardware storage requirements, the proposed algorithm can lead to ultra-low hardware resources and power consumption of hardware implementations, which is critical for the future multi-channel micro-systems. Both clinical and synthetic neural recording data sets were employed for evaluating the accuracy of the stream-based Hebbian eigenfilter. The performance of spike sorting using stream-based eigenfilter and the computational complexity of the eigenfilter were rigorously evaluated and compared with conventional PCA algorithms. Field programmable logic arrays (FPGAs) were employed to implement the proposed algorithm, evaluate the hardware implementations and demonstrate the reduction in both power consumption and hardware memories achieved by the streaming computing

Results and discussion

Results demonstrate that the stream-based eigenfilter can achieve the same accuracy and is 10 times more computationally efficient when compared with conventional PCA algorithms. Hardware evaluations show that 90.3% logic resources, 95.1% power consumption and 86.8% computing latency can be reduced by the stream-based eigenfilter when compared with PCA hardware. By utilizing the streaming method, 92% memory resources and 67% power consumption can be saved when compared with the direct implementation of GHA.

Conclusion

Stream-based Hebbian eigenfilter presents a novel approach to enable real-time spike sorting with reduced computational complexity and hardware costs. This new design can be further utilized for multi-channel neuro-physiological experiments or chronic implants.     

Normalization for probabilistic inference with neurons

Recently, there have been a number of proposals regarding how biologically plausible neural networks might perform probabilistic inference (Rao, Neural Computation, 16(1):1-38, 2004; Eliasmith and Anderson, Neural engineering: computation, representation and dynamics in neurobiological systems, 2003; Ma et?al., Nature Neuroscience, 9(11):1432-1438, 2006; Sahani and Dayan, Neural Computation, 15(10):2255-2279, 2003). To be able to repeatedly perform such inference, it is essential that the represented distributions be appropriately normalized. Past approaches have considered normalization mechanisms independently of inference, often leaving them unexplored, or appealing to a notion of divisive normalization that requires pooling across many neurons. Here, we demonstrate how normalization and inference can be combined into an appropriate connection matrix, eliminating the need for pooling or a division-like operation. We algebraically demonstrate that such a solution is available regardless of the inference being performed. We show that such a solution is relevant to neural computation by implementing it in a recurrent spiking neural network.     

Dendritic normalisation improves learning in sparsely connected artificial neural networks

Artificial neural networks, taking inspiration from biological neurons, have become an invaluable tool for machine learning applications. Recent studies have developed techniques to effectively tune the connectivity of sparsely-connected artificial neural networks, which have the potential to be more computationally efficient than their fully-connected counterparts and more closely resemble the architectures of biological systems. We here present a normalisation, based on the biophysical behaviour of neuronal dendrites receiving distributed synaptic inputs, that divides the weight of an artificial neuron’s afferent contacts by their number. We apply this dendritic normalisation to various sparsely-connected feedforward network architectures, as well as simple recurrent and self-organised networks with spatially extended units. The learning performance is significantly increased, providing an improvement over other widely-used normalisations in sparse networks. The results are two-fold, being both a practical advance in machine learning and an insight into how the structure of neuronal dendritic arbours may contribute to computation.     

Neural hierarchical models of ecological populations

Neural networks are increasingly being used in science to infer hidden dynamics of natural systems from noisy observations, a task typically handled by hierarchical models in ecology. This article describes a class of hierarchical models parameterised by neural networks – neural hierarchical models. The derivation of such models analogises the relationship between regression and neural networks. A case study is developed for a neural dynamic occupancy model of North American bird populations, trained on millions of detection/non‐detection time series for hundreds of species, providing insights into colonisation and extinction at a continental scale. Flexible models are increasingly needed that scale to large data and represent ecological processes. Neural hierarchical models satisfy this need, providing a bridge between deep learning and ecological modelling that combines the function representation power of neural networks with the inferential capacity of hierarchical models.     

A case for spiking neural network simulation based on configurable multiple-FPGA systems

Recent neuropsychological research has begun to reveal that neurons encode information in the timing of spikes. Spiking neural network simulations are a flexible and powerful method for investigating the behaviour of neuronal systems. Simulation of the spiking neural networks in software is unable to rapidly generate output spikes in large-scale of neural network. An alternative approach, hardware implementation of such system, provides the possibility to generate independent spikes precisely and simultaneously output spike waves in real time, under the premise that spiking neural network can take full advantage of hardware inherent parallelism. We introduce a configurable FPGA-oriented hardware platform for spiking neural network simulation in this work. We aim to use this platform to combine the speed of dedicated hardware with the programmability of software so that it might allow neuroscientists to put together sophisticated computation experiments of their own model. A feed-forward hierarchy network is developed as a case study to describe the operation of biological neural systems (such as orientation selectivity of visual cortex) and computational models of such systems. This model demonstrates how a feed-forward neural network constructs the circuitry required for orientation selectivity and provides platform for reaching a deeper understanding of the primate visual system. In the future, larger scale models based on this framework can be used to replicate the actual architecture in visual cortex, leading to more detailed predictions and insights into visual perception phenomenon.     

Event-driven simulation of neural population synchronization facilitated by electrical coupling

Most neural communication and processing tasks are driven by spikes. This has enabled the application of the event-driven simulation schemes. However the simulation of spiking neural networks based on complex models that cannot be simplified to analytical expressions (requiring numerical calculation) is very time consuming. Here we describe briefly an event-driven simulation scheme that uses pre-calculated table-based neuron characterizations to avoid numerical calculations during a network simulation, allowing the simulation of large-scale neural systems. More concretely we explain how electrical coupling can be simulated efficiently within this computation scheme, reproducing synchronization processes observed in detailed simulations of neural populations.     

Maximum Likelihood Neural Network Prediction Models

Neural networks have received much attention in recent years mostly by non-statisticians. The purpose of this paper is to incorporate neural networks in a non-linear regression model and obtain maximum likelihood estimates of the network parameters using a standard Newton-Raphson algorithm. We use maximum likelihood estimators instead of the usual back-propagation technique and compare the neural network predictions with predictions of quadratic regression models and with non-parametric nearest neighbor predictions. These comparisons are made using data generated from a variety of functions. Because of the number of parameters involved, neural network models can easily over-fit the data, hence validation of results is crucial.     

Fading memory and kernel properties of generic cortical microcircuit models

It is quite difficult to construct circuits of spiking neurons that can carry out complex computational tasks. On the other hand even randomly connected circuits of spiking neurons can in principle be used for complex computational tasks such as time-warp invariant speech recognition. This is possible because such circuits have an inherent tendency to integrate incoming information in such a way that simple linear readouts can be trained to transform the current circuit activity into the target output for a very large number of computational tasks. Consequently we propose to analyze circuits of spiking neurons in terms of their roles as analog fading memory and non-linear kernels, rather than as implementations of specific computational operations and algorithms. This article is a sequel to [W. Maass, T. Natschl?ger, H. Markram, Real-time computing without stable states: a new framework for neural computation based on perturbations, Neural Comput. 14 (11) (2002) 2531-2560, Online available as #130 from: ], and contains new results about the performance of generic neural microcircuit models for the recognition of speech that is subject to linear and non-linear time-warps, as well as for computations on time-varying firing rates. These computations rely, apart from general properties of generic neural microcircuit models, just on capabilities of simple linear readouts trained by linear regression. This article also provides detailed data on the fading memory property of generic neural microcircuit models, and a quick review of other new results on the computational power of such circuits of spiking neurons.     

A generalized locomotion CPG architecture based on oscillatory building blocks

 Neural oscillation is one of the most extensively investigated topics of artificial neural networks. Scientific approaches to the functionalities of both natural and artificial intelligences are strongly related to mechanisms underlying oscillatory activities. This paper concerns itself with the assumption of the existence of central pattern generators (CPGs), which are the plausible neural architectures with oscillatory capabilities, and presents a discrete and generalized approach to the functionality of locomotor CPGs of legged animals. Based on scheduling by multiple edge reversal (SMER), a primitive and deterministic distributed algorithm, it is shown how oscillatory building block (OBB) modules can be created and, hence, how OBB-based networks can be formulated as asymmetric Hopfield-like neural networks for the generation of complex coordinated rhythmic patterns observed among pairs of biological motor neurons working during different gait patterns. It is also shown that the resulting Hopfield-like network possesses the property of reproducing the whole spectrum of different gaits intrinsic to the target locomotor CPGs. Although the new approach is not restricted to the understanding of the neurolocomotor system of any particular animal, hexapodal and quadrupedal gait patterns are chosen as illustrations given the wide interest expressed by the ongoing research in the area. Received: 14 June 2002 / Accepted: 18 February 2003 / Published online: 20 May 2003 Correspondence to: Z. Yang (e-mail: zhijun.yang@ed.ac.uk) Acknowledgements. This work was partially supported by CNPq, the Brazilian Research Agency, under support number 143032/96-8. We are grateful for the helpful discussions with Prof. V.C. Barbosa, Dr. A.E. Xavier, Dr. M.S. Dutra, and Dr. A.F.R. Araújo. The donations of FPGA hardware and software from XILINX Incorporation under the order No. XUP2930 and XUP3576 are also highly appreciated.     

Dynamics of excitable neural networks with heterogeneous connectivity

A central issue of neuroscience is to understand how neural units integrates internal and external signals to create coherent states. Recently, it has been shown that the sensitivity and dynamic range of neural assemblies are optimal at a critical coupling among its elements. Complex architectures of connections seem to play a constructive role on the reliable coordination of neural units. Here we show that, the synchronizability and sensitivity of excitable neural networks can be tuned by diversity in the connections strengths. We illustrate our findings for weighted networks with regular, random and complex topologies. Additional comparisons of real brain networks support previous studies suggesting that heterogeneity in the connectivity may play a constructive role on information processing. These findings provide insights into the relationship between structure and function of neural circuits.