Alan B. Stokes

A Spiking Neural Network Model of the Lateral Geniculate Nucleus on the SpiNNaker Machine

We present a spiking neural network model of the thalamic Lateral Geniculate Nucleus (LGN) developed on SpiNNaker, which is a state-of-the-art digital neuromorphic hardware built with very-low-power ARM processors. The parallel, event-based data processing in SpiNNaker makes it viable for building massively parallel neuro-computational frameworks. The LGN model has 140 neurons representing a “basic building block” for larger modular architectures. The motivation of this work is to simulate biologically plausible LGN dynamics on SpiNNaker. Synaptic layout of the model is consistent with biology. The model response is validated with existing literature reporting entrainment in steady state visually evoked potentials (SSVEP)—brain oscillations corresponding to periodic visual stimuli recorded via electroencephalography (EEG). Periodic stimulus to the model is provided by: a synthetic spike-train with inter-spike-intervals in the range 10–50 Hz at a resolution of 1 Hz; and spike-train output from a state-of-the-art electronic retina subjected to a light emitting diode flashing at 10, 20, and 40 Hz, simulating real-world visual stimulus to the model. The resolution of simulation is 0.1 ms to ensure solution accuracy for the underlying differential equations defining Izhikevichs neuron model. Under this constraint, 1 s of model simulation time is executed in 10 s real time on SpiNNaker; this is because simulations on SpiNNaker work in real time for time-steps dt ⩾ 1 ms. The model output shows entrainment with both sets of input and contains harmonic components of the fundamental frequency. However, suppressing the feed-forward inhibition in the circuit produces subharmonics within the gamma band (>30 Hz) implying a reduced information transmission fidelity. These model predictions agree with recent lumped-parameter computational model-based predictions, using conventional computers. Scalability of the framework is demonstrated by a multi-node architecture consisting of three “nodes,” where each node is the “basic building block” LGN model. This 420 neuron model is tested with synthetic periodic stimulus at 10 Hz to all the nodes. The model output is the average of the outputs from all nodes, and conforms to the above-mentioned predictions of each node. Power consumption for model simulation on SpiNNaker is ≪1 W.

SensorBench: benchmarking approaches to processing wireless sensor network data

Wireless sensor networks enable cost-effective data collection for tasks such as precision agriculture and environment monitoring. However, the resource-constrained nature of sensor nodes, which often have both limited computational capabilities and battery lifetimes, means that applications that use them must make judicious use of these resources. Research that seeks to support data intensive sensor applications has explored a range of approaches and developed many different techniques, including bespoke algorithms for specific analyses and generic sensor network query processors. However, all such proposals sit within a multi-dimensional design space, where it can be difficult to understand the implications of specific decisions and to identify optimal solutions. This paper presents a benchmark that seeks to support the systematic analysis and comparison of different techniques and platforms, enabling both development and user communities to make well informed choices. The contributions of the paper include: (i) the identification of key variables and performance metrics; (ii) the specification of experiments that explore how different types of task perform under different metrics for the controlled variables; and (iii) an application of the benchmark to investigate the behavior of several representative platforms and techniques.

Adapting to node failure in sensor network query processing

The typical nodes used in mote-level wireless sensor networks (WSNs) are often brittle and severely resource-constrained. In particular, nodes are often battery-powered, thereby making energy depletion a significant risk. When changes to the connectivity graph occur as a result of node failure, the overall computation may collapse unless it is capable of adapting to the new WSN state. Sensor network query processors (SNQPs) construe a WSN as a distributed, continuous query platform where the streams of sensed values constitute the logical extents of interest. Crucially, in the context of this paper, they must make assumptions about the connectivity graph of the WSN at compile time that are likely not to hold for the lifetime of the compiled query evaluation plans (QEPs) the SNQPs generate. This paper address the problem of ensuring that a QEP continues to execute even if some nodes fail. The goal is to extend the lifetime of the QEP, i.e., the period during which it produces results, beyond the point where node failures start to occur. We contribute descriptions of two different approaches that have been implemented in an existing SNQP and present experimental results indicating that each significantly increases the overall lifetime of a query compared with non adaptive approach.

Resilient sensor network query processing using logical overlays

The typical nodes used in mote-level wireless sensor networks (WSNs) are often brittle and severely resource-constrained. In particular, nodes are often battery-powered, thereby making energy depletion a significant risk. When changes to the connectivity graph occur as a result of node failure, the overall computation may collapse unless it is capable of adapting to the new WSN state. Sensor network query processors (SNQPs) construe a WSN as a distributed, continuous query platform where the streams of sensed values constitute the logical extents of interest. Crucially, in the context of this paper, they must make assumptions about the connectivity graph of the WSN at compile time that are likely not to hold for the lifetime of the compiled query evaluation plan (QEP) the SNQPs generate. This paper addresses the problem of extending the lifetime of an evaluating QEP in the event of node failures. The basic idea is to derive an equivalence class over the nodes in the WSN that are equipotent for a given QEP and then to assign each QEP fragment instance to a set of equipotent nodes (rather than a single one). In this respect, the scheduling of QEP fragment instances is onto an overlay network of logical nodes, each of which maps to many physical nodes in the connectivity graph. We contribute a description of how this approach has been implemented in an existing SNQP and present experimental results indicating that it significantly increases the overall lifetime of a query whilst incurring small runtime adaptation costs.

Proactive adaptations in sensor network query processing

Wireless sensor networks (WSN) are used by many applications for event and environmental monitoring. Due to the resource-limited nodes in WSNs, there has been much research into extending the functional lifetime of the network through energy-saving techniques. Sensor Network Query Processing (SNQP) is one such technique. SNQP uses information about a query and the WSN over which it is to be run, to generate an energy-efficient Query Execution Plan (QEP) that distributes processing in the form of QEP fragments to the nodes in the WSN. However, any QEP is likely to drain the batteries of the nodes unevenly, and, as a result, nodes used in a QEP may run out of energy when there are significant energy stocks still available in the WSN. An adaptive query processor could react to energy depletion, for example, by generating a revised plan that refrains from using the drained nodes. However, adapting only when a node has been depleted may provide few opportunities for the creation of effective new QEPs. In this paper, we introduce an approach that determines, at query compilation time, a sequence of QEPs with switch times for transitioning between successive plans with a view to extending the overall lifetime of the query. We describe how this approach has been implemented as an extension to an existing SNQP and present experimental results indicating that it can significantly increase QEP lifetimes.

A Collaborative Simulation-Analysis Workflow for Computational Neuroscience Using HPC

Workflows for the acquisition and analysis of data in the natural sciences exhibit a growing degree of complexity and heterogeneity, are increasingly performed in large collaborative efforts, and often require the use of high-performance computing (HPC). Here, we explore the reasons for these new challenges and demands and discuss their impact with a focus on the scientific domain of computational neuroscience. We argue for the need of software platforms integrating HPC systems that allow scientists to construct, comprehend and execute workflows composed of diverse data generation and processing steps using different tools. As a use case we present a concrete implementation of such a complex workflow, covering diverse topics such as HPC-based simulation using the NEST software, access to the SpiNNaker neuromorphic hardware platform, complex data analysis using the Elephant library, and interactive visualization methods for facilitating further analysis. Tools are embedded into a web-based software platform under development by the Human Brain Project, called the Collaboratory. On the basis of this implementation, we discuss the state of the art and future challenges in constructing large, collaborative workflows with access to HPC resources.

Transport-Independent Protocols for Universal AER Communications

The emergence of Address-Event Representation (AER) as a general communications method across a large variety of neural devices suggests that they might be made interoperable. If there were a standard AER interface, systems could communicate using native AER signalling, allowing the construction of large-scale, real-time, heterogeneous neural systems. We propose a transport-agnostic AER protocol that permits direct bidirectional event communications between systems over Ethernet, and demonstrate practical implementations that connect a neuromimetic chip: SpiNNaker, both to standard host PCs and to real-time robotic systems. The protocol specifies a header and packet format that supports a variety of different possible packet types while coping with questions of data alignment, time sequencing, and packet compression. Such a model creates a flexible solution either for real-time communications between neural devices or for live spike I/O and visualisation in a host PC. With its standard physical layer and flexible protocol, the specification provides a prototype for AER protocol standardisation that is at once compatible with legacy systems and expressive enough for future very-large-scale neural systems.

Behavioral Learning in a Cognitive Neuromorphic Robot: An Integrative Approach

We present here a learning system using the iCub humanoid robot and the SpiNNaker neuromorphic chip to solve the real-world task of object-specific attention. Integrating spiking neural networks with robots introduces considerable complexity for questionable benefit if the objective is simply task performance. But, we suggest, in a cognitive robotics context, where the goal is understanding how to compute, such an approach may yield useful insights to neural architecture as well as learned behavior, especially if dedicated neural hardware is available. Recent advances in cognitive robotics and neuromorphic processing now make such systems possible. Using a scalable, structured, modular approach, we build a spiking neural network where the effects and impact of learning can be predicted and tested, and the network can be scaled or extended to new tasks automatically. We introduce several enhancements to a basic network and show how they can be used to direct performance toward behaviorally relevant goals. Results show that using a simple classical spike-timing-dependent plasticity (STDP) rule on selected connections, we can get the robot (and network) to progress from poor task-specific performance to good performance. Behaviorally relevant STDP appears to contribute strongly to positive learning: “do this” but less to negative learning: “don’t do that.” In addition, we observe that the effect of structural enhancements tends to be cumulative. The overall system suggests that it is by being able to exploit combinations of effects, rather than any one effect or property in isolation, that spiking networks can achieve compelling, task-relevant behavior.

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.