Corey M. Thibeault

HRLSim: A High Performance Spiking Neural Network Simulator for GPGPU Clusters

Modeling of large-scale spiking neural models is an important tool in the quest to understand brain function and subsequently create real-world applications. This paper describes a spiking neural network simulator environment called HRL Spiking Simulator (HRLSim). This simulator is suitable for implementation on a cluster of general purpose graphical processing units (GPGPUs). Novel aspects of HRLSim are described and an analysis of its performance is provided for various configurations of the cluster. With the advent of inexpensive GPGPU cards and compute power, HRLSim offers an affordable and scalable tool for design, real-time simulation, and analysis of large-scale spiking neural networks.

Using a hybrid neuron in physiologically inspired models of the basal ganglia

Our current understanding of the basal ganglia (BG) has facilitated the creation of computational models that have contributed novel theories, explored new functional anatomy and demonstrated results complementing physiological experiments. However, the utility of these models extends beyond these applications. Particularly in neuromorphic engineering, where the basal ganglias role in computation is important for applications such as power efficient autonomous agents and model-based control strategies. The neurons used in existing computational models of the BG, however, are not amenable for many low-power hardware implementations. Motivated by a need for more hardware accessible networks, we replicate four published models of the BG, spanning single neuron and small networks, replacing the more computationally expensive neuron models with an Izhikevich hybrid neuron. This begins with a network modeling action-selection, where the basal activity levels and the ability to appropriately select the most salient input is reproduced. A Parkinsons disease model is then explored under normal conditions, Parkinsonian conditions and during subthalamic nucleus deep brain stimulation (DBS). The resulting network is capable of replicating the loss of thalamic relay capabilities in the Parkinsonian state and its return under DBS. This is also demonstrated using a network capable of action-selection. Finally, a study of correlation transfer under different patterns of Parkinsonian activity is presented. These networks successfully captured the significant results of the originals studies. This not only creates a foundation for neuromorphic hardware implementations but may also support the development of large-scale biophysical models. The former potentially providing a way of improving the efficacy of DBS and the latter allowing for the efficient simulation of larger more comprehensive networks.

Efficiently passing messages in distributed spiking neural network simulation

Efficiently passing spiking messages in a neural model is an important aspect of high-performance simulation. As the scale of networks has increased so has the size of the computing systems required to simulate them. In addition, the information exchange of these resources has become more of an impediment to performance. In this paper we explore spike message passing using different mechanisms provided by the Message Passing Interface (MPI). A specific implementation, MVAPICH, designed for high-performance clusters with Infiniband hardware is employed. The focus is on providing information about these mechanisms for users of commodity high-performance spiking simulators. In addition, a novel hybrid method for spike exchange was implemented and benchmarked.

Modeling oxytocin induced neurorobotic trust and intent recognition in human-robot interaction

Recent human pharmacological fMRI studies suggest that oxytocin (OT) is a centrally-acting neurotransmitter important in the development and expression of trusting relationships in men and women. OT administration in humans was shown to increase trust, acceptance of social risk, memory of faces, and inference of the emotional state of others, in part by directly inhibiting the amygdala. However, the cerebral microcircuitry underlying this mechanism is still unclear. Here, we propose a spiking integrate-and-fire neuronal model of several key interacting brain regions affected by OT neurophysiology during social trust behavior. As a social behavior scenario, we embodied the brain simulator in a behaving virtual humanoid neurorobot, which interacted with a human via a camera. At the physiological level, the amygdala tonic firing was modeled using our recurrent asynchronous irregular nonlinear (RAIN) network architecture. OT cells were modeled with triple apical dendrites characteristic of their structure in the paraventricular nucleus of the hypothalamus. Our architecture demonstrated the success of our system in learning trust by discriminating concordant from discordant movements of a human actor. This led to a cooperative versus protective behavior by the neurorobot after being challenged by a new intent.

Analyzing large-scale spiking neural data with HRLAnalysis™

The additional capabilities provided by high-performance neural simulation environments and modern computing hardware has allowed for the modeling of increasingly larger spiking neural networks. This is important for exploring more anatomically detailed networks but the corresponding accumulation in data can make analyzing the results of these simulations difficult. This is further compounded by the fact that many existing analysis packages were not developed with large spiking data sets in mind. Presented here is a software suite developed to not only process the increased amount of spike-train data in a reasonable amount of time, but also provide a user friendly Python interface. We describe the design considerations, implementation and features of the HRLAnalysis™ suite. In addition, performance benchmarks demonstrating the speedup of this design compared to a published Python implementation are also presented. The result is a high-performance analysis toolkit that is not only usable and readily extensible, but also straightforward to interface with existing Python modules.

Breaking the virtual barrier: real-time interactions with spiking neural models

The unrivaled complexity of the human brain has driven many researchers towards larger and more detailed models of neural processing. Often run on remote high-performance computing architectures, these simulations can be difficult to access at significant levels of detail. In general, after simulations are completed, the results are analyzed off-line. This paradigm can make development of large models exploring complex and time-consuming simulations, such as learning or persistent neural activity, very difficult. Presented here is a toolkit, dubbed NCSTools, used for real-time interactions with large-scale neural simulations run on the NeoCortical Simulator (NCS). NCSTools is a remote monitoring package that provides a number of options for input, output, and modification of a running simulation. Input stimuli can be voltage, current, or probabilities of firing. For output there are several options for both collection and visualization. Information can be compiled as cell voltages, currents, synaptic efficacy, or spike-events within a population. The model information flow can be altered dynamically by NCSTools, as can model parameters such as spike timing dependent plasticity, long-term potentiation, and long-term depression. Additionally, NCSTools includes a simple network server interface that supports multiple connections from different client programs. These programs can be used for control of defined NCSTools actions, monitoring specific neural information, or synchronization with the simulation. This allows dynamic construction of distributed neural processing systems. Recently, NCSTools was utilized for real-time interactions between a Hebbian-STDP enabled neural simulation, a virtual robotic interface, and a human participant. Presented in Figure ​Figure11 is a diagrammatic view of the learning scenario. Processed visual information, in this case the color of a card, was sent to NCSTools. The information was then converted and sent to the neural simulation running on a remote computing cluster. As the simulation proceeded, the competing neural areas of visual and motor processing were monitored by NCSTools. The resulting activity was correlated with a pointing action to one of two colored balls that matched the color presented. After the robot completed pointing, NCSTools would provide the human interacting with the simulation an interface to reward the robot if it pointed to the correct ball. The reward, analogous to a dopaminergic increase, resulted in an STDP dependent increase in synaptic efficacy. Figure 1 Diagram of a Virtual Neuro-Robotic Interface constructed around NCSTools. The ability to monitor and modify simulations in real-time can be incredibly useful in large-scale spiking-network research. More importantly, this demonstrates another step towards multi-scale visualization of neural simulations in a virtual environment.

A role for neuromorphic processors in therapeutic nervous system stimulation

The motivations behind the development of many neuromorphic processors have been dominated by either the creation of better artificial intelligence, or novel non-von Neumann computing paradigms. A result of this impetus has been a number of low-power processors capable of simulating many different biological features of the nervous system. Power efficiency is crucial for deployed neuromorphic systems, but it also opens this technology up to other energy restricted applications. In this opinion, we suggest two such applications pertaining to therapeutic stimulation of the nervous system where closing the control loop could be assisted by advances in neuromorphic architectures: (1) deep brain stimulation (DBS) in the treatment of Parkinsons disease and (2) epidural spinal cord stimulation (ESS) for restoring voluntary motor functions. Though there are still questions that must be addressed before this would be feasible, but we are suggesting that the technological barriers—in both the algorithms and hardware—can be overcome with directed funding and research. Neuromorphic processor research is centered around the creation of brain-like intelligence through power-efficient circuits that borrow elements directly from biology (Mead, 1989). The applications for these projects range from brain-scale simulations (Gao et al., 2012; Benjamin et al., 2014) and in silica experimentation (Schemmel et al., 2010; Furber et al., 2012), to brain-like computing and learning (Merolla et al., 2011; Srinivasa and Cruz-Albrecht, 2012; Cruz-Albrecht et al., 2013; Rahimi Azghadi et al., 2014; Schmuker et al., 2014). These projects promise unrivaled access to large-scale models of the brain as well as insight into the unique non-von Neumann computation that biological systems appear to achieve. Regardless of the motivation, the tangible result of these efforts has been an accumulation of low-power circuits capable of emulating various elements of the nervous system. Although these are essential for embodying robotic systems and augmenting current super-computing paradigms, they also have the potential to assist in nervous system stimulation control. This application is outside the scope of the currently funded neuromorphic hardware projects, but with new insights and technological advances, it is one that will be particularly beneficial.

Reward-based learning for virtual neurorobotics through emotional speech processing

Reward-based learning can easily be applied to real life with a prevalence in children teaching methods. It also allows machines and software agents to automatically determine the ideal behavior from a simple reward feedback (e.g., encouragement) to maximize their performance. Advancements in affective computing, especially emotional speech processing (ESP) have allowed for more natural interaction between humans and robots. Our research focuses on integrating a novel ESP system in a relevant virtual neurorobotic (VNR) application. We created an emotional speech classifier that successfully distinguished happy and utterances. The accuracy of the system was 95.3 and 98.7% during the offline mode (using an emotional speech database) and the live mode (using live recordings), respectively. It was then integrated in a neurorobotic scenario, where a virtual neurorobot had to learn a simple exercise through reward-based learning. If the correct decision was made the robot received a spoken reward, which in turn stimulated synapses (in our simulated model) undergoing spike-timing dependent plasticity (STDP) and reinforced the corresponding neural pathways. Both our ESP and neurorobotic systems allowed our neurorobot to successfully and consistently learn the exercise. The integration of ESP in real-time computational neuroscience architecture is a first step toward the combination of human emotions and virtual neurorobotics.

Algorithm for Reliable Detection of Beat Onsets in Cerebral Blood Flow Velocity Signals

Transcranial Doppler (TCD) ultrasound has been demonstrated to be a valuable tool for assessing cerebral hemodynamics via measurement of cerebral blood flow velocity (CBFV), with a number of established clinical indications. However, CBFV waveform analysis depends on reliable pulse onset detection, an inherently difficult task for CBFV signals acquired via TCD. We study the application of a new algorithm for CBFV pulse segmentation, which locates pulse onsets in a sequential manner using a moving difference filter and adaptive thresholding. The test data set used in this study consists of 92,012 annotated CBFV pulses, whose quality is representative of real world data. On this test set, the algorithm achieves a true positive rate of 99.998% (2 false negatives), positive predictive value of 99.998% (2 false positives), and mean temporal offset error of 6.10 ± 4.75 ms. We do note that in this context, the way in which true positives, false positives, and false negatives are defined caries some nuance, so care should be taken when drawing comparisons to other algorithms. Additionally, we find that 97.8% and 99.5% of onsets are detected within 10 ms and 30 ms, respectively, of the true onsets. The algorithm’s performance in spite of the large degree of variation in signal quality and waveform morphology present in the test data suggests that it may serve as a valuable tool for the accurate and reliable identification of CBFV pulse onsets in neurocritical care settings.

Velocity Curvature Index: a Novel Diagnostic Biomarker for Large Vessel Occlusion

Despite being a conveniently portable technology for stroke assessment, Transcranial Doppler ultrasound (TCD) remains widely underutilized due to complex training requirements necessary to reliably obtain and interpret cerebral blood flow velocity (CBFV) waveforms. The validation of objective TCD metrics for large vessel occlusion (LVO) represents a first critical step toward enabling use by less formally trained personnel. In this work, we assess the diagnostic utility, relative to current standard CT angiography (CTA), of a novel TCD-derived biomarker for detecting LVO. Patients admitted to the hospital with stroke symptoms underwent TCD screening and were grouped into LVO and control groups based on the presence of CTA confirmed occlusion. Velocity curvature index (VCI) was computed from CBFV waveforms recorded at multiple depths from the middle cerebral arteries (MCA) of both cerebral hemispheres. VCI was assessed for 66 patients, 33 of which had occlusions of the MCA or internal carotid artery. Our results show that VCI was more informative when measured from the cerebral hemisphere ipsilateral to the site of occlusion relative to contralateral. Moreover, given any pair of bilateral recordings, VCI separated LVO patients from controls with average area under receiver operating characteristic curve of 92%, which improved to greater than 94% when pairs were selected by maximal velocity. We conclude that VCI is an analytically valid candidate biomarker for LVO diagnosis, possessing comparable accuracy, and several important advantages, relative to current TCD diagnostic methodologies.