Is a 4-bit synaptic weight resolution enough? - constraints on enabling spike-timing dependent plasticity in neuromorphic hardware

Adaptive STDP-based on-chip spike pattern detection

A spiking neural network (SNN) is a bottom-up tool used to describe information processing in brain microcircuits. It is becoming a crucial neuromorphic computational model. Spike-timing-dependent plasticity (STDP) is an unsupervised brain-like learning rule implemented in many SNNs and neuromorphic chips. However, a significant performance gap exists between ideal model simulation and neuromorphic implementation. The performance of STDP learning in neuromorphic chips deteriorates because the resolution of synaptic efficacy in such chips is generally restricted to 6 bits or less, whereas simulations employ the entire 64-bit floating-point precision available on digital computers. Previously, we introduced a bio-inspired learning rule named adaptive STDP and demonstrated via numerical simulation that adaptive STDP (using only 4-bit fixed-point synaptic efficacy) performs similarly to STDP learning (using 64-bit floating-point precision) in a noisy spike pattern detection model. Herein, we present the experimental results demonstrating the performance of adaptive STDP learning. To the best of our knowledge, this is the first study that demonstrates unsupervised noisy spatiotemporal spike pattern detection to perform well and maintain the simulation performance on a mixed-signal CMOS neuromorphic chip with low-resolution synaptic efficacy. The chip was designed in Taiwan Semiconductor Manufacturing Company (TSMC) 250 nm CMOS technology node and comprises a soma circuit and 256 synapse circuits along with their learning circuitry.

A System-on-Chip Based Hybrid Neuromorphic Compute Node Architecture for Reproducible Hyper-Real-Time Simulations of Spiking Neural Networks

Despite the great strides neuroscience has made in recent decades, the underlying principles of brain function remain largely unknown. Advancing the field strongly depends on the ability to study large-scale neural networks and perform complex simulations. In this context, simulations in hyper-real-time are of high interest, as they would enable both comprehensive parameter scans and the study of slow processes, such as learning and long-term memory. Not even the fastest supercomputer available today is able to meet the challenge of accurate and reproducible simulation with hyper-real acceleration. The development of novel neuromorphic computer architectures holds out promise, but the high costs and long development cycles for application-specific hardware solutions makes it difficult to keep pace with the rapid developments in neuroscience. However, advances in System-on-Chip (SoC) device technology and tools are now providing interesting new design possibilities for application-specific implementations. Here, we present a novel hybrid software-hardware architecture approach for a neuromorphic compute node intended to work in a multi-node cluster configuration. The node design builds on the Xilinx Zynq-7000 SoC device architecture that combines a powerful programmable logic gate array (FPGA) and a dual-core ARM Cortex-A9 processor extension on a single chip. Our proposed architecture makes use of both and takes advantage of their tight coupling. We show that available SoC device technology can be used to build smaller neuromorphic computing clusters that enable hyper-real-time simulation of networks consisting of tens of thousands of neurons, and are thus capable of meeting the high demands for modeling and simulation in neuroscience.

Dynamical Characteristics of Recurrent Neuronal Networks Are Robust Against Low Synaptic Weight Resolution

The representation of the natural-density, heterogeneous connectivity of neuronal network models at relevant spatial scales remains a challenge for Computational Neuroscience and Neuromorphic Computing. In particular, the memory demands imposed by the vast number of synapses in brain-scale network simulations constitute a major obstacle. Limiting the number resolution of synaptic weights appears to be a natural strategy to reduce memory and compute load. In this study, we investigate the effects of a limited synaptic-weight resolution on the dynamics of recurrent spiking neuronal networks resembling local cortical circuits and develop strategies for minimizing deviations from the dynamics of networks with high-resolution synaptic weights. We mimic the effect of a limited synaptic weight resolution by replacing normally distributed synaptic weights with weights drawn from a discrete distribution, and compare the resulting statistics characterizing firing rates, spike-train irregularity, and correlation coefficients with the reference solution. We show that a naive discretization of synaptic weights generally leads to a distortion of the spike-train statistics. If the weights are discretized such that the mean and the variance of the total synaptic input currents are preserved, the firing statistics remain unaffected for the types of networks considered in this study. For networks with sufficiently heterogeneous in-degrees, the firing statistics can be preserved even if all synaptic weights are replaced by the mean of the weight distribution. We conclude that even for simple networks with non-plastic neurons and synapses, a discretization of synaptic weights can lead to substantial deviations in the firing statistics unless the discretization is performed with care and guided by a rigorous validation process. For the network model used in this study, the synaptic weights can be replaced by low-resolution weights without affecting its macroscopic dynamical characteristics, thereby saving substantial amounts of memory.

An Adaptive STDP Learning Rule for Neuromorphic Systems

The promise of neuromorphic computing to develop ultra-low-power intelligent devices lies in its ability to localize information processing and memory storage in synaptic circuits much like the synapses in the brain. Spiking neural networks modeled using high-resolution synapses and armed with local unsupervised learning rules like spike time-dependent plasticity (STDP) have shown promising results in tasks such as pattern detection and image classification. However, designing and implementing a conventional, multibit STDP circuit becomes complex both in terms of the circuitry and the required silicon area. In this work, we introduce a modified and hardware-friendly STDP learning (named adaptive STDP) implemented using just 4-bit synapses. We demonstrate the capability of this learning rule in a pattern recognition task, in which a neuron learns to recognize a specific spike pattern embedded within noisy inhomogeneous Poisson spikes. Our results demonstrate that the performance of the proposed learning rule (94% using just 4-bit synapses) is similar to the conventional STDP learning (96% using 64-bit floating-point precision). The models used in this study are ideal ones for a CMOS neuromorphic circuit with analog soma and synapse circuits and mixed-signal learning circuits. The learning circuit stores the synaptic weight in a 4-bit digital memory that is updated asynchronously. In circuit simulation with Taiwan Semiconductor Manufacturing Company (TSMC) 250 nm CMOS process design kit (PDK), the static power consumption of a single synapse and the energy per spike (to generate a synaptic current of amplitude 15 pA and time constant 3 ms) are less than 2 pW and 200 fJ, respectively. The static power consumption of the learning circuit is less than 135 pW, and the energy to process a pair of pre- and postsynaptic spikes corresponding to a single learning step is less than 235 pJ. A single 4-bit synapse (capable of being configured as excitatory, inhibitory, or shunting inhibitory) along with its learning circuitry and digital memory occupies around 17,250 μm² of silicon area.

Considerations for Neuromorphic Supercomputing in Semiconducting and Superconducting Optoelectronic Hardware

Any large-scale spiking neuromorphic system striving for complexity at the level of the human brain and beyond will need to be co-optimized for communication and computation. Such reasoning leads to the proposal for optoelectronic neuromorphic platforms that leverage the complementary properties of optics and electronics. Starting from the conjecture that future large-scale neuromorphic systems will utilize integrated photonics and fiber optics for communication in conjunction with analog electronics for computation, we consider two possible paths toward achieving this vision. The first is a semiconductor platform based on analog CMOS circuits and waveguide-integrated photodiodes. The second is a superconducting approach that utilizes Josephson junctions and waveguide-integrated superconducting single-photon detectors. We discuss available devices, assess scaling potential, and provide a list of key metrics and demonstrations for each platform. Both platforms hold potential, but their development will diverge in important respects. Semiconductor systems benefit from a robust fabrication ecosystem and can build on extensive progress made in purely electronic neuromorphic computing but will require III-V light source integration with electronics at an unprecedented scale, further advances in ultra-low capacitance photodiodes, and success from emerging memory technologies. Superconducting systems place near theoretically minimum burdens on light sources (a tremendous boon to one of the most speculative aspects of either platform) and provide new opportunities for integrated, high-endurance synaptic memory. However, superconducting optoelectronic systems will also contend with interfacing low-voltage electronic circuits to semiconductor light sources, the serial biasing of superconducting devices on an unprecedented scale, a less mature fabrication ecosystem, and cryogenic infrastructure.

In situ Parallel Training of Analog Neural Network Using Electrochemical Random-Access Memory

In-memory computing based on non-volatile resistive memory can significantly improve the energy efficiency of artificial neural networks. However, accurate in situ training has been challenging due to the nonlinear and stochastic switching of the resistive memory elements. One promising analog memory is the electrochemical random-access memory (ECRAM), also known as the redox transistor. Its low write currents and linear switching properties across hundreds of analog states enable accurate and massively parallel updates of a full crossbar array, which yield rapid and energy-efficient training. While simulations predict that ECRAM based neural networks achieve high training accuracy at significantly higher energy efficiency than digital implementations, these predictions have not been experimentally achieved. In this work, we train a 3 × 3 array of ECRAM devices that learns to discriminate several elementary logic gates (AND, OR, NAND). We record the evolution of the network's synaptic weights during parallel in situ (on-line) training, with outer product updates. Due to linear and reproducible device switching characteristics, our crossbar simulations not only accurately simulate the epochs to convergence, but also quantitatively capture the evolution of weights in individual devices. The implementation of the first in situ parallel training together with strong agreement with simulation results provides a significant advance toward developing ECRAM into larger crossbar arrays for artificial neural network accelerators, which could enable orders of magnitude improvements in energy efficiency of deep neural networks.

Neurons with dendrites can perform linearly separable computations with low resolution synaptic weights

In theory, neurons modelled as single layer perceptrons can implement all linearly separable computations. In practice, however, these computations may require arbitrarily precise synaptic weights. This is a strong constraint since both biological neurons and their artificial counterparts have to cope with limited precision. Here, we explore how non-linear processing in dendrites helps overcome this constraint. We start by finding a class of computations which requires increasing precision with the number of inputs in a perceptron and show that it can be implemented without this constraint in a neuron with sub-linear dendritic subunits. Then, we complement this analytical study by a simulation of a biophysical neuron model with two passive dendrites and a soma, and show that it can implement this computation. This work demonstrates a new role of dendrites in neural computation: by distributing the computation across independent subunits, the same computation can be performed more efficiently with less precise tuning of the synaptic weights. This work not only offers new insight into the importance of dendrites for biological neurons, but also paves the way for new, more efficient architectures of artificial neuromorphic chips.

Fast and Accurate Sparse Coding of Visual Stimuli With a Simple, Ultralow-Energy Spiking Architecture

Memristive crossbars have become a popular means for realizing unsupervised and supervised learning techniques. In previous neuromorphic architectures with leaky integrate-and-fire neurons, the crossbar itself has been separated from the neuron capacitors to preserve mathematical rigor. In this paper, we sought to design a simplified sparse coding circuit without this restriction, resulting in a fast circuit that approximated a sparse coding operation at a minimal loss in accuracy. We showed that connecting the neurons directly to the crossbar resulted in a more energy-efficient sparse coding architecture and alleviated the need to prenormalize receptive fields. This paper provides derivations for the design of such a network, named the simple spiking locally competitive algorithm, as well as CMOS designs and results on the CIFAR and MNIST data sets. Compared to a nonspiking, nonapproximate model which scored 33% on CIFAR-10 with a single-layer classifier, this hardware scored 32% accuracy. When used with a state-of-the-art deep learning classifier, the nonspiking model achieved 82% and our simplified, spiking model achieved 80% while compressing the input data by 92%. Compared to a previously proposed spiking model, our proposed hardware consumed 99% less energy to do the same work at 21 × the throughput. Accuracy held out with online learning to a write variance of 3%, suitable for the often reported 4-bit resolution required for neuromorphic algorithms, with offline learning to a write variance of 27%, and with read variance to 40%. The proposed architecture's excellent accuracy, throughput, and significantly lower energy usage demonstrate the utility of our innovations.

Deep Learning With Spiking Neurons: Opportunities and Challenges

Spiking neural networks (SNNs) are inspired by information processing in biology, where sparse and asynchronous binary signals are communicated and processed in a massively parallel fashion. SNNs on neuromorphic hardware exhibit favorable properties such as low power consumption, fast inference, and event-driven information processing. This makes them interesting candidates for the efficient implementation of deep neural networks, the method of choice for many machine learning tasks. In this review, we address the opportunities that deep spiking networks offer and investigate in detail the challenges associated with training SNNs in a way that makes them competitive with conventional deep learning, but simultaneously allows for efficient mapping to hardware. A wide range of training methods for SNNs is presented, ranging from the conversion of conventional deep networks into SNNs, constrained training before conversion, spiking variants of backpropagation, and biologically motivated variants of STDP. The goal of our review is to define a categorization of SNN training methods, and summarize their advantages and drawbacks. We further discuss relationships between SNNs and binary networks, which are becoming popular for efficient digital hardware implementation. Neuromorphic hardware platforms have great potential to enable deep spiking networks in real-world applications. We compare the suitability of various neuromorphic systems that have been developed over the past years, and investigate potential use cases. Neuromorphic approaches and conventional machine learning should not be considered simply two solutions to the same classes of problems, instead it is possible to identify and exploit their task-specific advantages. Deep SNNs offer great opportunities to work with new types of event-based sensors, exploit temporal codes and local on-chip learning, and we have so far just scratched the surface of realizing these advantages in practical applications.

Feedback control for microring weight banks

Microring weight banks present novel opportunities for reconfigurable, high-performance analog signal processing in photonics. Controlling microring filter response is a challenge due to fabrication variations and thermal sensitivity. Prior work showed continuous weight control of multiple wavelength-division multiplexed signals in a bank of microrings based on calibration and feedforward control. Other prior work has shown resonance locking based on feedback control by monitoring photoabsorption-induced changes in resistance across in-ring photoconductive heaters. In this work, we demonstrate continuous, multi-channel control of a microring weight bank with an effective 5.1 bits of accuracy on 2Gbps signals. Unlike resonance locking, the approach relies on an estimate of filter transmission versus photo-induced resistance changes. We introduce an estimate still capable of providing 4.2 bits of accuracy without any direct transmission measurements. Furthermore, we present a detailed characterization of this response for different values of carrier wavelength offset and power. Feedback weight control renders tractable the weight control problem in reconfigurable analog photonic networks.

Neural and Synaptic Array Transceiver: A Brain-Inspired Computing Framework for Embedded Learning

Embedded, continual learning for autonomous and adaptive behavior is a key application of neuromorphic hardware. However, neuromorphic implementations of embedded learning at large scales that are both flexible and efficient have been hindered by a lack of a suitable algorithmic framework. As a result, most neuromorphic hardware are trained off-line on large clusters of dedicated processors or GPUs and transferred post hoc to the device. We address this by introducing the neural and synaptic array transceiver (NSAT), a neuromorphic computational framework facilitating flexible and efficient embedded learning by matching algorithmic requirements and neural and synaptic dynamics. NSAT supports event-driven supervised, unsupervised and reinforcement learning algorithms including deep learning. We demonstrate the NSAT in a wide range of tasks, including the simulation of Mihalas-Niebur neuron, dynamic neural fields, event-driven random back-propagation for event-based deep learning, event-based contrastive divergence for unsupervised learning, and voltage-based learning rules for sequence learning. We anticipate that this contribution will establish the foundation for a new generation of devices enabling adaptive mobile systems, wearable devices, and robots with data-driven autonomy.

Data and Power Efficient Intelligence with Neuromorphic Learning Machines

The success of deep networks and recent industry involvement in brain-inspired computing is igniting a widespread interest in neuromorphic hardware that emulates the biological processes of the brain on an electronic substrate. This review explores interdisciplinary approaches anchored in machine learning theory that enable the applicability of neuromorphic technologies to real-world, human-centric tasks. We find that (1) recent work in binary deep networks and approximate gradient descent learning are strikingly compatible with a neuromorphic substrate; (2) where real-time adaptability and autonomy are necessary, neuromorphic technologies can achieve significant advantages over main-stream ones; and (3) challenges in memory technologies, compounded by a tradition of bottom-up approaches in the field, block the road to major breakthroughs. We suggest that a neuromorphic learning framework, tuned specifically for the spatial and temporal constraints of the neuromorphic substrate, will help guiding hardware algorithm co-design and deploying neuromorphic hardware for proactive learning of real-world data.

Neuromorphic photonic networks using silicon photonic weight banks

Photonic systems for high-performance information processing have attracted renewed interest. Neuromorphic silicon photonics has the potential to integrate processing functions that vastly exceed the capabilities of electronics. We report first observations of a recurrent silicon photonic neural network, in which connections are configured by microring weight banks. A mathematical isomorphism between the silicon photonic circuit and a continuous neural network model is demonstrated through dynamical bifurcation analysis. Exploiting this isomorphism, a simulated 24-node silicon photonic neural network is programmed using “neural compiler” to solve a differential system emulation task. A 294-fold acceleration against a conventional benchmark is predicted. We also propose and derive power consumption analysis for modulator-class neurons that, as opposed to laser-class neurons, are compatible with silicon photonic platforms. At increased scale, Neuromorphic silicon photonics could access new regimes of ultrafast information processing for radio, control, and scientific computing.

Demonstrating Hybrid Learning in a Flexible Neuromorphic Hardware System

We present results from a new approach to learning and plasticity in neuromorphic hardware systems: to enable flexibility in implementable learning mechanisms while keeping high efficiency associated with neuromorphic implementations, we combine a general-purpose processor with full-custom analog elements. This processor is operating in parallel with a fully parallel neuromorphic system consisting of an array of synapses connected to analog, continuous time neuron circuits. Novel analog correlation sensor circuits process spike events for each synapse in parallel and in real-time. The processor uses this pre-processing to compute new weights possibly using additional information following its program. Therefore, to a certain extent, learning rules can be defined in software giving a large degree of flexibility. Synapses realize correlation detection geared towards Spike-Timing Dependent Plasticity (STDP) as central computational primitive in the analog domain. Operating at a speed-up factor of 1000 compared to biological time-scale, we measure time-constants from tens to hundreds of micro-seconds. We analyze variability across multiple chips and demonstrate learning using a multiplicative STDP rule. We conclude that the presented approach will enable flexible and efficient learning as a platform for neuroscientific research and technological applications.

Stochastic Synapses Enable Efficient Brain-Inspired Learning Machines

Recent studies have shown that synaptic unreliability is a robust and sufficient mechanism for inducing the stochasticity observed in cortex. Here, we introduce Synaptic Sampling Machines (S2Ms), a class of neural network models that uses synaptic stochasticity as a means to Monte Carlo sampling and unsupervised learning. Similar to the original formulation of Boltzmann machines, these models can be viewed as a stochastic counterpart of Hopfield networks, but where stochasticity is induced by a random mask over the connections. Synaptic stochasticity plays the dual role of an efficient mechanism for sampling, and a regularizer during learning akin to DropConnect. A local synaptic plasticity rule implementing an event-driven form of contrastive divergence enables the learning of generative models in an on-line fashion. S2Ms perform equally well using discrete-timed artificial units (as in Hopfield networks) or continuous-timed leaky integrate and fire neurons. The learned representations are remarkably sparse and robust to reductions in bit precision and synapse pruning: removal of more than 75% of the weakest connections followed by cursory re-learning causes a negligible performance loss on benchmark classification tasks. The spiking neuron-based S2Ms outperform existing spike-based unsupervised learners, while potentially offering substantial advantages in terms of power and complexity, and are thus promising models for on-line learning in brain-inspired hardware.

Comparing Neuromorphic Solutions in Action: Implementing a Bio-Inspired Solution to a Benchmark Classification Task on Three Parallel-Computing Platforms

Neuromorphic computing employs models of neuronal circuits to solve computing problems. Neuromorphic hardware systems are now becoming more widely available and "neuromorphic algorithms" are being developed. As they are maturing toward deployment in general research environments, it becomes important to assess and compare them in the context of the applications they are meant to solve. This should encompass not just task performance, but also ease of implementation, speed of processing, scalability, and power efficiency. Here, we report our practical experience of implementing a bio-inspired, spiking network for multivariate classification on three different platforms: the hybrid digital/analog Spikey system, the digital spike-based SpiNNaker system, and GeNN, a meta-compiler for parallel GPU hardware. We assess performance using a standard hand-written digit classification task. We found that whilst a different implementation approach was required for each platform, classification performances remained in line. This suggests that all three implementations were able to exercise the model's ability to solve the task rather than exposing inherent platform limits, although differences emerged when capacity was approached. With respect to execution speed and power consumption, we found that for each platform a large fraction of the computing time was spent outside of the neuromorphic device, on the host machine. Time was spent in a range of combinations of preparing the model, encoding suitable input spiking data, shifting data, and decoding spike-encoded results. This is also where a large proportion of the total power was consumed, most markedly for the SpiNNaker and Spikey systems. We conclude that the simulation efficiency advantage of the assessed specialized hardware systems is easily lost in excessive host-device communication, or non-neuronal parts of the computation. These results emphasize the need to optimize the host-device communication architecture for scalability, maximum throughput, and minimum latency. Moreover, our results indicate that special attention should be paid to minimize host-device communication when designing and implementing networks for efficient neuromorphic computing.

Robustness of spiking Deep Belief Networks to noise and reduced bit precision of neuro-inspired hardware platforms

Increasingly large deep learning architectures, such as Deep Belief Networks (DBNs) are the focus of current machine learning research and achieve state-of-the-art results in different domains. However, both training and execution of large-scale Deep Networks require vast computing resources, leading to high power requirements and communication overheads. The on-going work on design and construction of spike-based hardware platforms offers an alternative for running deep neural networks with significantly lower power consumption, but has to overcome hardware limitations in terms of noise and limited weight precision, as well as noise inherent in the sensor signal. This article investigates how such hardware constraints impact the performance of spiking neural network implementations of DBNs. In particular, the influence of limited bit precision during execution and training, and the impact of silicon mismatch in the synaptic weight parameters of custom hybrid VLSI implementations is studied. Furthermore, the network performance of spiking DBNs is characterized with regard to noise in the spiking input signal. Our results demonstrate that spiking DBNs can tolerate very low levels of hardware bit precision down to almost two bits, and show that their performance can be improved by at least 30% through an adapted training mechanism that takes the bit precision of the target platform into account. Spiking DBNs thus present an important use-case for large-scale hybrid analog-digital or digital neuromorphic platforms such as SpiNNaker, which can execute large but precision-constrained deep networks in real time.

A neuromorphic implementation of multiple spike-timing synaptic plasticity rules for large-scale neural networks

We present a neuromorphic implementation of multiple synaptic plasticity learning rules, which include both Spike Timing Dependent Plasticity (STDP) and Spike Timing Dependent Delay Plasticity (STDDP). We present a fully digital implementation as well as a mixed-signal implementation, both of which use a novel dynamic-assignment time-multiplexing approach and support up to 2²⁶ (64M) synaptic plasticity elements. Rather than implementing dedicated synapses for particular types of synaptic plasticity, we implemented a more generic synaptic plasticity adaptor array that is separate from the neurons in the neural network. Each adaptor performs synaptic plasticity according to the arrival times of the pre- and post-synaptic spikes assigned to it, and sends out a weighted or delayed pre-synaptic spike to the post-synaptic neuron in the neural network. This strategy provides great flexibility for building complex large-scale neural networks, as a neural network can be configured for multiple synaptic plasticity rules without changing its structure. We validate the proposed neuromorphic implementations with measurement results and illustrate that the circuits are capable of performing both STDP and STDDP. We argue that it is practical to scale the work presented here up to 2³⁶ (64G) synaptic adaptors on a current high-end FPGA platform.

Turn Down That Noise: Synaptic Encoding of Afferent SNR in a Single Spiking Neuron

We have added a simplified neuromorphic model of Spike Time Dependent Plasticity (STDP) to the previously described Synapto-dendritic Kernel Adapting Neuron (SKAN), a hardware efficient neuron model capable of learning spatio-temporal spike patterns. The resulting neuron model is the first to perform synaptic encoding of afferent signal-to-noise ratio in addition to the unsupervised learning of spatio-temporal spike patterns. The neuron model is particularly suitable for implementation in digital neuromorphic hardware as it does not use any complex mathematical operations and uses a novel shift-based normalization approach to achieve synaptic homeostasis. The neuron's noise compensation properties are characterized and tested on random spatio-temporal spike patterns as well as a noise corrupted subset of the zero images of the MNIST handwritten digit dataset. Results show the simultaneously learning common patterns in its input data while dynamically weighing individual afferents based on their signal to noise ratio. Despite its simplicity the interesting behaviors of the neuron model and the resulting computational power may also offer insights into biological systems.

A framework for plasticity implementation on the SpiNNaker neural architecture

Many of the precise biological mechanisms of synaptic plasticity remain elusive, but simulations of neural networks have greatly enhanced our understanding of how specific global functions arise from the massively parallel computation of neurons and local Hebbian or spike-timing dependent plasticity rules. For simulating large portions of neural tissue, this has created an increasingly strong need for large scale simulations of plastic neural networks on special purpose hardware platforms, because synaptic transmissions and updates are badly matched to computing style supported by current architectures. Because of the great diversity of biological plasticity phenomena and the corresponding diversity of models, there is a great need for testing various hypotheses about plasticity before committing to one hardware implementation. Here we present a novel framework for investigating different plasticity approaches on the SpiNNaker distributed digital neural simulation platform. The key innovation of the proposed architecture is to exploit the reconfigurability of the ARM processors inside SpiNNaker, dedicating a subset of them exclusively to process synaptic plasticity updates, while the rest perform the usual neural and synaptic simulations. We demonstrate the flexibility of the proposed approach by showing the implementation of a variety of spike- and rate-based learning rules, including standard Spike-Timing dependent plasticity (STDP), voltage-dependent STDP, and the rate-based BCM rule. We analyze their performance and validate them by running classical learning experiments in real time on a 4-chip SpiNNaker board. The result is an efficient, modular, flexible and scalable framework, which provides a valuable tool for the fast and easy exploration of learning models of very different kinds on the parallel and reconfigurable SpiNNaker system.

Racing to learn: statistical inference and learning in a single spiking neuron with adaptive kernels

This paper describes the Synapto-dendritic Kernel Adapting Neuron (SKAN), a simple spiking neuron model that performs statistical inference and unsupervised learning of spatiotemporal spike patterns. SKAN is the first proposed neuron model to investigate the effects of dynamic synapto-dendritic kernels and demonstrate their computational power even at the single neuron scale. The rule-set defining the neuron is simple: there are no complex mathematical operations such as normalization, exponentiation or even multiplication. The functionalities of SKAN emerge from the real-time interaction of simple additive and binary processes. Like a biological neuron, SKAN is robust to signal and parameter noise, and can utilize both in its operations. At the network scale neurons are locked in a race with each other with the fastest neuron to spike effectively "hiding" its learnt pattern from its neighbors. The robustness to noise, high speed, and simple building blocks not only make SKAN an interesting neuron model in computational neuroscience, but also make it ideal for implementation in digital and analog neuromorphic systems which is demonstrated through an implementation in a Field Programmable Gate Array (FPGA). Matlab, Python, and Verilog implementations of SKAN are available at: http://www.uws.edu.au/bioelectronics_neuroscience/bens/reproducible_research.

Characterization and compensation of network-level anomalies in mixed-signal neuromorphic modeling platforms

Advancing the size and complexity of neural network models leads to an ever increasing demand for computational resources for their simulation. Neuromorphic devices offer a number of advantages over conventional computing architectures, such as high emulation speed or low power consumption, but this usually comes at the price of reduced configurability and precision. In this article, we investigate the consequences of several such factors that are common to neuromorphic devices, more specifically limited hardware resources, limited parameter configurability and parameter variations due to fixed-pattern noise and trial-to-trial variability. Our final aim is to provide an array of methods for coping with such inevitable distortion mechanisms. As a platform for testing our proposed strategies, we use an executable system specification (ESS) of the BrainScaleS neuromorphic system, which has been designed as a universal emulation back-end for neuroscientific modeling. We address the most essential limitations of this device in detail and study their effects on three prototypical benchmark network models within a well-defined, systematic workflow. For each network model, we start by defining quantifiable functionality measures by which we then assess the effects of typical hardware-specific distortion mechanisms, both in idealized software simulations and on the ESS. For those effects that cause unacceptable deviations from the original network dynamics, we suggest generic compensation mechanisms and demonstrate their effectiveness. Both the suggested workflow and the investigated compensation mechanisms are largely back-end independent and do not require additional hardware configurability beyond the one required to emulate the benchmark networks in the first place. We hereby provide a generic methodological environment for configurable neuromorphic devices that are targeted at emulating large-scale, functional neural networks.

An event-based neural network architecture with an asynchronous programmable synaptic memory

We present a hybrid analog/digital very large scale integration (VLSI) implementation of a spiking neural network with programmable synaptic weights. The synaptic weight values are stored in an asynchronous Static Random Access Memory (SRAM) module, which is interfaced to a fast current-mode event-driven DAC for producing synaptic currents with the appropriate amplitude values. These currents are further integrated by current-mode integrator synapses to produce biophysically realistic temporal dynamics. The synapse output currents are then integrated by compact and efficient integrate and fire silicon neuron circuits with spike-frequency adaptation and adjustable refractory period and spike-reset voltage settings. The fabricated chip comprises a total of 32 × 32 SRAM cells, 4 × 32 synapse circuits and 32 × 1 silicon neurons. It acts as a transceiver, receiving asynchronous events in input, performing neural computation with hybrid analog/digital circuits on the input spikes, and eventually producing digital asynchronous events in output. Input, output, and synaptic weight values are transmitted to/from the chip using a common communication protocol based on the Address Event Representation (AER). Using this representation it is possible to interface the device to a workstation or a micro-controller and explore the effect of different types of Spike-Timing Dependent Plasticity (STDP) learning algorithms for updating the synaptic weights values in the SRAM module. We present experimental results demonstrating the correct operation of all the circuits present on the chip.

A neuromorphic network for generic multivariate data classification

Computational neuroscience has uncovered a number of computational principles used by nervous systems. At the same time, neuromorphic hardware has matured to a state where fast silicon implementations of complex neural networks have become feasible. En route to future technical applications of neuromorphic computing the current challenge lies in the identification and implementation of functional brain algorithms. Taking inspiration from the olfactory system of insects, we constructed a spiking neural network for the classification of multivariate data, a common problem in signal and data analysis. In this model, real-valued multivariate data are converted into spike trains using "virtual receptors" (VRs). Their output is processed by lateral inhibition and drives a winner-take-all circuit that supports supervised learning. VRs are conveniently implemented in software, whereas the lateral inhibition and classification stages run on accelerated neuromorphic hardware. When trained and tested on real-world datasets, we find that the classification performance is on par with a naïve Bayes classifier. An analysis of the network dynamics shows that stable decisions in output neuron populations are reached within less than 100 ms of biological time, matching the time-to-decision reported for the insect nervous system. Through leveraging a population code, the network tolerates the variability of neuronal transfer functions and trial-to-trial variation that is inevitably present on the hardware system. Our work provides a proof of principle for the successful implementation of a functional spiking neural network on a configurable neuromorphic hardware system that can readily be applied to real-world computing problems.

Pattern classification by memristive crossbar circuits using ex situ and in situ training

Memristors are memory resistors that promise the efficient implementation of synaptic weights in artificial neural networks. Whereas demonstrations of the synaptic operation of memristors already exist, the implementation of even simple networks is more challenging and has yet to be reported. Here we demonstrate pattern classification using a single-layer perceptron network implemented with a memrisitive crossbar circuit and trained using the perceptron learning rule by ex situ and in situ methods. In the first case, synaptic weights, which are realized as conductances of titanium dioxide memristors, are calculated on a precursor software-based network and then imported sequentially into the crossbar circuit. In the second case, training is implemented in situ, so the weights are adjusted in parallel. Both methods work satisfactorily despite significant variations in the switching behaviour of the memristors. These results give hope for the anticipated efficient implementation of artificial neuromorphic networks and pave the way for dense, high-performance information processing systems.

Reward-based learning under hardware constraints-using a RISC processor embedded in a neuromorphic substrate

In this study, we propose and analyze in simulations a new, highly flexible method of implementing synaptic plasticity in a wafer-scale, accelerated neuromorphic hardware system. The study focuses on globally modulated STDP, as a special use-case of this method. Flexibility is achieved by embedding a general-purpose processor dedicated to plasticity into the wafer. To evaluate the suitability of the proposed system, we use a reward modulated STDP rule in a spike train learning task. A single layer of neurons is trained to fire at specific points in time with only the reward as feedback. This model is simulated to measure its performance, i.e., the increase in received reward after learning. Using this performance as baseline, we then simulate the model with various constraints imposed by the proposed implementation and compare the performance. The simulated constraints include discretized synaptic weights, a restricted interface between analog synapses and embedded processor, and mismatch of analog circuits. We find that probabilistic updates can increase the performance of low-resolution weights, a simple interface between analog synapses and processor is sufficient for learning, and performance is insensitive to mismatch. Further, we consider communication latency between wafer and the conventional control computer system that is simulating the environment. This latency increases the delay, with which the reward is sent to the embedded processor. Because of the time continuous operation of the analog synapses, delay can cause a deviation of the updates as compared to the not delayed situation. We find that for highly accelerated systems latency has to be kept to a minimum. This study demonstrates the suitability of the proposed implementation to emulate the selected reward modulated STDP learning rule. It is therefore an ideal candidate for implementation in an upgraded version of the wafer-scale system developed within the BrainScaleS project.

Six networks on a universal neuromorphic computing substrate

In this study, we present a highly configurable neuromorphic computing substrate and use it for emulating several types of neural networks. At the heart of this system lies a mixed-signal chip, with analog implementations of neurons and synapses and digital transmission of action potentials. Major advantages of this emulation device, which has been explicitly designed as a universal neural network emulator, are its inherent parallelism and high acceleration factor compared to conventional computers. Its configurability allows the realization of almost arbitrary network topologies and the use of widely varied neuronal and synaptic parameters. Fixed-pattern noise inherent to analog circuitry is reduced by calibration routines. An integrated development environment allows neuroscientists to operate the device without any prior knowledge of neuromorphic circuit design. As a showcase for the capabilities of the system, we describe the successful emulation of six different neural networks which cover a broad spectrum of both structure and functionality.

An FPGA Implementation of a Polychronous Spiking Neural Network with Delay Adaptation

We present an FPGA implementation of a re-configurable, polychronous spiking neural network with a large capacity for spatial-temporal patterns. The proposed neural network generates delay paths de novo, so that only connections that actually appear in the training patterns will be created. This allows the proposed network to use all the axons (variables) to store information. Spike Timing Dependent Delay Plasticity is used to fine-tune and add dynamics to the network. We use a time multiplexing approach allowing us to achieve 4096 (4k) neurons and up to 1.15 million programmable delay axons on a Virtex 6 FPGA. Test results show that the proposed neural network is capable of successfully recalling more than 95% of all spikes for 96% of the stored patterns. The tests also show that the neural network is robust to noise from random input spikes.