Search

Dissertation, RWTH Aachen University, 2023; Jülich : Forschungszentrum Jülich GmbH, Zentralbibliothek, Verlag 1 Online-Ressource : Illustrationen, Diagramme (2023). = Dissertation, RWTH Aachen University, 2023, The brain uses intricate biological mechanisms and principles to solve a variety of tasks. These principles endow systems with self-learning capabilities, efficient energy usage, and high storage capacity. A core concept that lies at the heart of brain computation is sequence prediction and replay. This form of computation is essential for almost all our daily tasks such as movement generation, perception, and language. Understanding how the brain performs such a computation advances neuroscience and paves the way for new technological brain-inspired applications. In the first part of this thesis, we propose a sequence learning model that explains how biological networks learn to predict upcoming elements, signal non-anticipated events, and recall sequences in response to a cue signal. The model accounts for anatomical and electrophysiological properties of cortical neuronal circuits, and learns complex sequences in an unsupervised manner by means of known biological plasticity and homeostatic control mechanisms. After learning, it self-organizes into a configuration characterized by a high degree of sparsity in connectivity and activity allowing for both high storage capacity and efficient energy usage. In the second part, we extend the sequence learning model such that it permits probabilistic sequential memory recall in response to ambiguous cues. In the absence of noise, the model deterministically recalls the sequence shown most frequently during training. We investigate how different forms of noise give rise to more exploratory behavior. We show that uncorrelated noise averages out in population based encoding leading to non exploratory dynamics. Locally coherent noise in the form of random stimulus locking to spatiotemporal oscillations addresses this issue. Our results show that depending on the amplitude and frequency of oscillation, the network can recall learned sequences according to different strategies: either always replay the most frequent sequence, or replay sequences according to their occurrence probability during training. The study contributes to an understanding of the neuronal mechanisms underlying different decision strategies in the face of ambiguity, and highlights the role of coherent network activity during sequential memory recall. Finally, we investigate the feasibility of implementing the sequence learning model on dedicated hardware mimicking brain properties. Here, we focus on a type of hardware where synapses are emulated by memristive devices. As a first step in this direction, we replace the synapse dynamics of the original model with dynamics describing the phenomenological behavior of memristive elements, and demonstrate resilience with respect to different device characteristics. In this thesis, we further describe how the sequence learning model can be adapted at the algorithmic level to foster an implementation in a full electronic circuit centered around a memristive crossbar array. Overall, this thesis sheds light on the key mechanisms underlying sequence learning, prediction, and replay in biological networks and demonstrates the feasibility of implementing this type of computation on neuromorphic hardware., Published by Forschungszentrum Jülich GmbH, Zentralbibliothek, Verlag, Jülich

The brain uses intricate biological mechanisms and principles to solve a variety of tasks. These principles endow systems with self-learning capabilities, efficient energy usage, and high storage capacity. A core concept that lies at the heart of brain computation is sequence prediction and replay. This form of computation is essential for almost all our daily tasks such as movement generation, perception, and language. Understanding how the brain performs such a computation advances neuroscience and paves the way for new technological brain-inspired applications. In the first part of this thesis, we propose a sequence learning model that explains how biological networks learn to predict upcoming elements, signal non-anticipated events, and recall sequences in response to a cue signal. The model accounts for anatomical and electrophysiological properties of cortical neuronal circuits, and learns complex sequences in an unsupervised manner by means of known biological plasticity and homeostatic control mechanisms. After learning, it self-organizes into a configuration characterized by a high degree of sparsity in connectivity and activity allowing for both high storage capacity and efficient energy usage. In the second part, we extend the sequence learning model such that it permits probabilistic sequential memory recall in response to ambiguous cues. In the absence of noise, the model deterministically recalls the sequence shown most frequently during training. We investigate how different forms of noise give rise to more exploratory behavior. We show that uncorrelated noise averages out in population based encoding leading to non-exploratory dynamics. Locally coherent noise in the form of random stimulus locking to spatiotemporal oscillations addresses this issue. Our results show that depending on the amplitude and frequency of oscillation, the network can recall learned sequences according to different strategies: either always replay the most frequent sequence, or replay sequences according to their occurrence probability during training. The study contributes to an understanding of the neuronal mechanisms underlying different decision strategies in the face of ambiguity, and highlights the role of coherent network activity during sequential memory recall. Finally, we investigate the feasibility of implementing the sequence learning model on dedicated hardware mimicking brain properties. Here, we focus on a type of hardware where synapses are emulated by memristive devices. As a first step in this direction, we replace the synapse dynamics of the original model with dynamics describing the phenomenological behavior of memristive elements, and demonstrate resilience with respect to different device characteristics. In this thesis, we further describe how the sequence learning model can be adapted at the algorithmic level to foster an implementation in a full electronic circuit centered around a memristive crossbar array. Overall, this thesis sheds light on the key mechanisms underlying sequence learning, prediction, and replay in biological networks and demonstrates the feasibility of implementing this type of computation on neuromorphic hardware.

Evolution gave humans advanced grasping capabili- ties combining an adaptive hand with efficient control. Grasping motions can quickly be adapted if the object moves or deforms. Soft-grasping with an anthropomorphic hand is a great capability for robots interacting with objects shaped for humans. Neverthe- less, most robotic applications use vacuum, 2-finger or custom made grippers. We present a biologically inspired spiking neural network (SNN) for soft-grasping to control a robotic hand. Two control loops are combined, one from motor primitives and one from a compliant controller activated by a reflex. The finger primitives represent synergies between joints and hand primitives represent different affordances. Contact is detected with a mech- anism based on inter-neuron circuits in the spinal cord to trigger reflexes. A Schunk SVH 5-finger hand was used to grasp objects with different shapes, stiffness and sizes. The SNN adapted the grasping motions without knowing the exact properties of the objects. The compliant controller with online learning proved to be sensitive, allowing even the grasping of balloons. In contrast to deep learning approaches, our SNN requires one example of each grasping motion to train the primitives. Computation of the inverse kinematics or complex contact point planning is not required. This approach simplifies the control and can be used on different robots providing similar adaptive features as a human hand. A physical imitation of a biological system implemented completely with SNN and a robotic hand can provide new insights into grasping mechanisms.

In recent years, neuroscientists have discovered that the neural information is encoded by spike trains with precise times. Supervised learning algorithm based on the precise times for spiking neurons becomes an important research field. Although many existing algorithms have the excellent learning ability, most of their mechanisms still have some complex computations and certain limitations. Moreover, the discontinuity of spiking process also makes it very difficult to build an efficient algorithm. This paper proposes a supervised learning algorithm for spiking neurons using the kernel function of spike trains based on a unit of pair-spike. Firstly, we comprehensively divide the intervals of spike trains. Then, we construct an optimal selection and computation method of spikes based on the unit of pair-spike. This method avoids some wrong computations and reduces the computational cost by using each effective input spike only once in every epoch. Finally, we use the kernel function defined by an inner product operator to solve the computing problem of discontinue spike process and multiple output spikes. The proposed algorithm is successfully applied to many learning tasks of spike trains, where the effect of our optimal selection and computation method is verified and the influence of learning factors such as learning kernel, learning rate, and learning epoch is analyzed. Moreover, compared with other algorithms, all experimental results show that our proposed algorithm has the higher learning accuracy and good learning efficiency.

How does reliable computation emerge from networks of noisy neurons? While individual neurons are intrinsically noisy, the collective dynamics of populations of neurons taken as a whole can be almost deterministic, supporting the hypothesis that, in the brain, computation takes place at the level of neuronal populations. Mathematical models of networks of noisy spiking neurons allow us to study the effects of neuronal noise on the dynamics of large networks. Classical mean-field models, i.e., models where all neurons are identical and where each neuron receives the average spike activity of the other neurons, offer toy examples where neuronal noise is absorbed in large networks, that is, large networks behave like deterministic systems. In particular, the dynamics of these large networks can be described by deterministic neuronal population equations. In this thesis, I first generalize classical mean-field limit proofs to a broad class of spiking neuron models that can exhibit spike-frequency adaptation and short-term synaptic plasticity, in addition to refractoriness. The mean-field limit can be exactly described by a multidimensional partial differential equation; the long time behavior of which can be rigorously studied using deterministic methods. Then, we show that there is a conceptual link between mean-field models for networks of spiking neurons and latent variable models used for the analysis of multi-neuronal recordings. More specifically, we use a recently proposed finite-size neuronal population equation, which we first mathematically clarify, to design a tractable Expectation-Maximization-type algorithm capable of inferring the latent population activities of multi-population spiking neural networks from the spike activity of a few visible neurons only, illustrating the idea that latent variable models can be seen as partially observed mean-field models. In classical mean-field models, neurons in large networks behave like independent, identically distributed processes driven by the average population activity -- a deterministic quantity, by the law of large numbers. The fact the neurons are identically distributed processes implies a form of redundancy that has not been observed in the cortex and which seems biologically implausible. To show, numerically, that the redundancy present in classical mean-field models is unnecessary for neuronal noise absorption in large networks, I construct a disordered network model where networks of spiking neurons behave like deterministic rate networks, despite the absence of redundancy. This last result suggests that the concentration of measure phenomenon, which generalizes the ``law of large numbers'' of classical mean-field models, might be an instrumental principle for understanding the emergence of noise-robust population dynamics in large networks of noisy neurons.

Artificial spiking neural networks are gaining increasing prominence due to their potential advantages over traditional, time-static artificial neural networks. Custom hardware implementations of spiking neural networks present many advantages over other implementation mediums. Two main topics are the focus of this work. Firstly, digital hardware implementations of spiking neurons and neuromorphic hardware are explored and presented. These implementations include novel implementations for lowered digital hardware requirements and reduced power consumption. The second section of this work proposes a novel method for selectively adding sparsity to a spiking neural network based on training set images for pattern recognition applications, thereby greatly reducing the inference time required in a digital hardware implementation.

This paper demonstrates a framework that entails a bottom-up approach to accelerate research, development, and verification of neuro-inspired sensing devices for real-life applications. Previous work in neuromorphic engineering mostly considered application-specific designs which is a strong limitation for researchers to develop novel applications and emulate the true behaviour of neuro-inspired systems. Hence to enable the fully parallel brain-like computations, this paper proposes a methodology where a spiking neuron model was emulated in software and electronic circuits were then implemented and characterized. The proposed approach offers a unique perspective whereby experimental measurements taken from a fabricated device allowing empirical models to be developed. This technique acts as a bridge between the theoretical and practical aspects of neuro-inspired devices. It is shown through software simulations and empirical modelling that the proposed technique is capable of replicating neural dynamics and post-synaptic potentials. Retrospectively, the proposed framework offers a first step towards open-source neuro-inspired hardware for a range of applications such as healthcare, applied machine learning and the internet of things (IoT).

Memory traces and associations between them are fundamental for cognitive brain function. Neuron recordings suggest that distributed assemblies of neurons in the brain serve as memory traces for spatial information, real-world items, and concepts. However, there is conflicting evidence regarding neural codes for associated memory traces. Some studies suggest the emergence of overlaps between assemblies during an association, while others suggest that the assemblies themselves remain largely unchanged and new assemblies emerge as neural codes for associated memory items. Here we study the emergence of neural codes for associated memory items in a generic computational model of recurrent networks of spiking neurons with a data-constrained rule for spike-timing-dependent plasticity. The model depends critically on 2 parameters, which control the excitability of neurons and the scale of initial synaptic weights. By modifying these 2 parameters, the model can reproduce both experimental data from the human brain on the fast formation of associations through emergent overlaps between assemblies, and rodent data where new neurons are recruited to encode the associated memories. Hence, our findings suggest that the brain can use both of these 2 neural codes for associations, and dynamically switch between them during consolidation.

In contrast with our everyday experience using brain circuits, it can take a prohibitively long time to train a computational system to produce the correct sequence of outputs in the presence of a series of inputs. This suggests that something important is missing in the way in which models are trying to reproduce basic cognitive functions. In this work, we introduce a new neuronal network architecture that is able to learn, in a single trial, an arbitrary long sequence of any known objects. The key point of the model is the explicit use of mechanisms and circuitry observed in the hippocampus, which allow the model to reach a level of efficiency and accuracy that, to the best of our knowledge, is not possible with abstract network implementations. By directly following the natural system’s layout and circuitry, this type of implementation has the additional advantage that the results can be more easily compared to experimental data, allowing a deeper and more direct understanding of the mechanisms underlying cognitive functions and dysfunctions, and opening the way to a new generation of learning architectures.

Empirical evidence suggests that children with autism spectrum disorder (ASD) show abnormal behavior during delay eyeblink conditioning. They show a higher conditioned response learning rate and earlier peak latency of the conditioned response signal. The neuronal mechanisms underlying this autistic behavioral phenotype are still unclear. Here, we use a physiologically constrained spiking neuron model of the cerebellar-cortical system to investigate which features are critical to explaining atypical learning in ASD. Significantly, the computer simulations run with the model suggest that the higher conditioned responses learning rate mainly depends on the reduced number of Purkinje cells. In contrast, the earlier peak latency mainly depends on the hyper-connections of the cerebellum with sensory and motor cortex. Notably, the model has been validated by reproducing the behavioral data collected from studies with real children. Overall, this article is a starting point to understanding the link between the behavioral and neurobiological basis in ASD learning. At the end of the paper, we discuss how this knowledge could be critical for devising new treatments.

Surface electromyogram (sEMG) signals result from muscle movement and hence they are an ideal candidate for benchmarking event-driven sensing and computing. We propose a simple yet novel approach for optimizing the spike encoding algorithm's hyper-parameters inspired by the readout layer concept in reservoir computing. Using a simple machine learning algorithm after spike encoding, we report performance higher than the state-of-the-art spiking neural networks on two open-source datasets for hand gesture recognition. The spike encoded data is processed through a spiking reservoir with a biologically inspired topology and neuron model. When trained with the unsupervised activity regulation CRITICAL algorithm to operate at the edge of chaos, the reservoir yields better performance than state-of-the-art convolutional neural networks. The reservoir performance with regulated activity was found to be 89.72% for the Roshambo EMG dataset and 70.6% for the EMG subset of sensor fusion dataset. Therefore, the biologically-inspired computing paradigm, which is known for being power efficient, also proves to have a great potential when compared with conventional AI algorithms., Comment: Accepted to International Conference on Neuromorphic Systems (ICONS 2021)

We address the problem of efficiently and informatively quantifying how multiplets of variables carry information about the future of the dynamical system they belong to. In particular we want to identify groups of variables carrying redundant or synergistic information, and track how the size and the composition of these multiplets changes as the collective behavior of the system evolves. In order to afford a parsimonious expansion of shared information, and at the same time control for lagged interactions and common effect, we develop a dynamical, conditioned version of the O-information, a framework recently proposed to quantify high-order interdependencies via multivariate extension of the mutual information. The dynamic O-information, here introduced, allows to separate multiplets of variables which influence synergistically the future of the system from redundant multiplets. We apply this framework to a dataset of spiking neurons from a monkey performing a perceptual discrimination task. The method identifies synergistic multiplets that include neurons previously categorized as containing little relevant information individually.

In this paper, we propose a novel approach for implementing spiking neurons through an optical system. Spiking neurons are a new approach to emulate the neural processes that occur in the brain, known as the third generation of artificial neural networks. They have been successfully used to build a new technology called neuromorphic engineering, which looks for a better performance than traditional computing in tasks usually performed by AI-based systems. Our optical system consists of a low-cost laser source, based on a microcontroller and a continuous-wave laser diode, the microcontroller allows producing synchronous or asynchronous pulses with complex time profiles. Here, through said system we have successfully reproduced most of the neural dynamics observed in biological neurons. These dynamics have been reproduced using a very simple optical array with a great potential for the development of neuromorphic systems. The optical system has been experimentally validated.

Combining information from multiple sources is a fundamental operation performed by networks of neurons in the brain, whose general principles are still largely unknown. Experimental evidence suggests that combination of inputs in cortex relies on nonlinear summation. Such nonlinearities are thought to be fundamental to perform complex computations. However, these non-linearities are inconsistent with the balanced-state model, one of the most popular models of cortical dynamics, which predicts networks have a linear response. This linearity is obtained in the limit of very large recurrent coupling strength. We investigate the stationary response of networks of spiking neurons as a function of coupling strength. We show that, while a linear transfer function emerges at strong coupling, nonlinearities are prominent at finite coupling, both at response onset and close to saturation. We derive a general framework to classify nonlinear responses in these networks and discuss which of them can be captured by rate models. This framework could help to understand the diversity of non-linearities observed in cortical networks., Author summary Models of cortical networks are often studied in the strong coupling limit, where the so-called balanced state emerges. Across a wide range of parameters, balanced state models explain a number of ubiquitous properties of cortex, such as irregular neural firing. However, in the strong coupling limit, balanced state models show an unrealistic linear network transfer function. We examined, in networks of spiking neurons, how nonlinearities arise as network coupling strength is reduced to realistic levels. We examine closed-form solutions that arise from mean-field analysis, and confirm results with numerical simulations, to show that nonlinearities at response onset and saturation emerge as coupling strength is reduced. Critically, for realistic parameter values, both types of nonlinearities are observed at experimentally-observed firing rates. Thus, cortical network models with moderate coupling strength can account for experimentally observed cortical response nonlinearities.

We propose a mean-field model of interacting point processes where each process has a memory of the time elapsed since its last event (age) and its recent past (leaky memory), generalizing Age-dependent Hawkes processes. The model is motivated by interacting nonlinear Hawkes processes with Markovian self-interaction and networks of spiking neurons with adaptation and short-term synaptic plasticity. By proving propagation of chaos and using a path integral representation for the law of the limit process, we show that, in the mean-field limit, the empirical measure of the system follows a multidimensional nonlocal transport equation., Comment: 22 pages

Low power/area cytomorphic chips may be interfaced and ultimately implanted in the human body for cell‐sensing and cell‐control applications of the future. In such electronic platforms, it is crucial to accurately mimic the biological timescales and operate in real time. This paper proposes a methodology where slow nonlinear dynamical systems describing the behaviour of naturally encountered biological systems can be efficiently realised in hardware. To this end, as a case study, a low power and efficient digital Application‐Specific Integrated Circuit capable of emulating slow intracellular calcium dynamics with timescales reaching to seconds has been fabricated in the commercially available AMS 0.35 μm technology and compared with its analog counterpart. The fabricated chip occupies an area of 1.5 mm2 (excluding the area of the pads) and consumes 18.93 nW for each calcium unit from a power supply of 3.3 V. The presented cytomimetic topology follows closely the behaviour of its biological counterpart, exhibiting similar time‐domain calcium ions dynamics. Results show that the implemented design has the potential to speed up large–scale simulations of slow intracellular dynamics by sharing cellular units in real time.

This paper studies dynamics of the bifurcating neuron circuit. Repeating integrate-and-fire behavior between a constant threshold and piecewise linear periodic base signal, the circuit can exhibit a variety of chaotic and super-stable periodic spike-trains. Deriving one-dimensional map of spike phases, the circuit dynamics can be analyzed precisely. Especially, this paper analyzes a variety of super-stable periodic spike-trains and steady state, transient phenomena. Analysis of the circuit is important not only as the basic study of nonlinear dynamical systems, but also for engineering applications.

Simulation of large-scale nonlinear dynamical systems on hardware with a high resemblance to their mathematical equivalents has been always a challenge in engineering. This paper presents a novel current-input current-output circuit supporting a systematic synthesis procedure of log-domain circuits capable of computing bilateral dynamical systems with considerably low power consumption and acceptable precision. Here, the application of the method is demonstrated by synthesizing four different case studies: 1) a relatively complex 2-D nonlinear neuron model; 2) a chaotic 3-D nonlinear dynamical system Lorenz attractor having arbitrary solutions for certain parameters; 3) a 2-D nonlinear Hopf oscillator, including bistability phenomenon sensitive to initial values; and 4) three small neurosynaptic networks comprising three FHN neuron models variously coupled with excitatory and inhibitory synapses. The validity of our approach is verified by nominal and Monte Carlo simulated results with realistic process parameters from the commercially available AMS 0.35- $\mu \text{m}$ technology. The resulting continuous-time, continuous-value, and low-power circuits exhibit various bifurcation phenomena, nominal time-domain responses in good agreement with their mathematical counterparts and fairly acceptable process variation results (less than 5% STD).

Recent studies have demonstrated that calcium is a widespread intracellular ion that controls a wide range of temporal dynamics in the mammalian body. The simulation and validation of such studies using experimental data would benefit from a fast large scale simulation and modelling tool. This paper presents a compact and fully reconfigurable cellular calcium model capable of mimicking Hopf bifurcation phenomenon and various nonlinear responses of the biological calcium dynamics. The proposed cellular model is synthesized on a digital platform for a single unit and a network model. Hardware synthesis, physical implementation on FPGA, and theoretical analysis confirm that the proposed cellular model can mimic the biological calcium behaviors with considerably low hardware overhead. The approach has the potential to speed up large-scale simulations of slow intracellular dynamics by sharing more cellular units in real-time. To this end, various networks constructed by pipelining 10 k to 40 k cellular calcium units are compared with an equivalent simulation run on a standard PC workstation. Results show that the cellular hardware model is, on average, 83 times faster than the CPU version.

International audience; In this paper, we present the synchronous approach as a new method for modeling neural architectures. The synchronous approach is initially used to design controllers for reactive real-time systems, and we explain how it fits well to our objective to design bio-inspired neuromorphic circuits for biohybrid experiments. We describe synchronous implementations of some neural models and validate our approach with the Brian simulator. We also compare the synchronous hardware code generated to a standard manually optimised code. And finally we discuss the next steps of our work.

Although the occurrence of Parkinsonian akinesia and tremor is traditionally associated to dopaminergic degeneration, the multifaceted neural processes that cause these impairments are not fully understood. As a consequence, current dopamine medications cannot be tailored to the specific dysfunctions of patients with the result that generic drug therapies produce different effects on akinesia and tremor. This article proposes a computational model focusing on the role of dopamine impairments in the occurrence of akinesia and resting tremor. The model has three key features, to date never integrated in a single computational system: (a) an architecture constrained on the basis of the relevant known system-level anatomy of the basal ganglia-thalamo-cortical loops; (b) spiking neurons with physiologically-constrained parameters; (c) a detailed simulation of the effects of both phasic and tonic dopamine release. The model exhibits a neural dynamics compatible with that recorded in the brain of primates and humans. Moreover, it suggests that akinesia might involve both tonic and phasic dopamine dysregulations whereas resting tremor might be primarily caused by impairments involving tonic dopamine release and the responsiveness of dopamine receptors. These results could lead to develop new therapies based on a system-level view of the Parkinson's disease and targeting phasic and tonic dopamine in differential ways.

The massively parallel nature of biological information processing plays an important role due to its superiority in comparison to human-engineered computing devices. In particular, it may hold the key to overcoming the von Neumann bottleneck that limits contemporary computer architectures. Physical-model neuromorphic devices seek to replicate not only this inherent parallelism, but also aspects of its microscopic dynamics in analog circuits emulating neurons and synapses. However, these machines require network models that are not only adept at solving particular tasks, but that can also cope with the inherent imperfections of analog substrates. We present a spiking network model that performs Bayesian inference through sampling on the BrainScaleS neuromorphic platform, where we use it for generative and discriminative computations on visual data. By illustrating its functionality on this platform, we implicitly demonstrate its robustness to various substrate-specific distortive effects, as well as its accelerated capability for computation. These results showcase the advantages of brain-inspired physical computation and provide important building blocks for large-scale neuromorphic applications.

A prominent feature of the dynamics of large neuronal networks are the synchrony-driven collective oscillations generated by the interplay between synaptic coupling and synaptic delays. This thesis investigates the emergence of delay-induced oscillations in networks of heterogeneous spiking neurons. Building on recent theoretical advances in exact mean field reductions for neuronal networks, this work explores the dynamics and bifurcations of an exact firing rate model with various forms of synaptic delays. In parallel, the results obtained using the novel firing rate model are compared with extensive numerical simulations of large networks of spiking neurons, which confirm the existence of numerous synchrony-based oscillatory states. Some of these states are novel and display complex forms of partial synchronization and collective chaos. Given the well-known limitation of traditional firing rate models to describe synchrony-based oscillations, previous studies greatly overlooked many of the oscillatory states found here. Therefore, this thesis provides a unique exploration of the oscillatory scenarios found in neuronal networks due to the presence of delays, and may substantially extend the mathematical tools available for modeling the plethora of oscillations detected in electrical recordings of brain activity. Una característica fonamental de la dinàmica d'una xarxa neuronal és l'emergència d'oscil·lacions degudes a sincronització. L'origen d'aquestes oscil·lacions és molt sovint degut les interaccions sinàptiques i als seus retards temporals inherents. Aquesta tesi analitza la emergència d'oscil·lacions produïdes per retards sinàptics en xarxes neuronals heterogènies. A partir de troballes recents en teories de camp mig per xarxes neuronals, aquest treball explora la dinàmica i les bifurcacions d'un model de {\it rate} amb diferents tipus de retards sinàptics. En paral·lel els resultats obtinguts mitjançant el nou model de rate són comparats amb simulacions numèriques de grans xarxes neuronals. Aquestes simulacions confirmen l'existència de nombrosos estats oscil·latoris produïts per sincronització. Alguns d'aquests estats són nous I mostren formes complexes de sincronització parcial i de caos col·lectiu. Gran part d'aquestes oscil·lacions han estat àmpliament ignorades a la literatura, degut a la limitació dels models tradicionals de rate per descriure estats amb un alt nivell de sincronització. Així doncs aquesta tesi ofereix una exploració única dels possibles escenaris oscil·latoris en xarxes neuronals amb retards sinàptics, i amplia significativament les eines matemàtiques disponibles per a la modelització de la gran diversitat d'oscil·lacions neuronals presents en les mesures elèctriques de l'activitat cerebral.

The sustainability of ever more sophisticated artificial intelligence relies on the continual development of highly energy‐efficient and compact computing hardware that mimics the biological neural networks. Recently, the neural firing properties have been widely explored in various spiking neuron devices, which could emerge as the fundamental building blocks of future neuromorphic/in‐memory computing hardware. By leveraging the intrinsic device characteristics, the device‐based spiking neuron has the potential advantage of a compact circuit area for implementing neural networks with high density and high parallelism. However, a comprehensive benchmark that considers not only the device but also the peripheral circuit necessary for realizing complete neural functions is still lacking. Herein, the recent progress of emerging spiking neuron devices and circuits is reviewed. By implementing peripheral analog circuits for supporting various spiking neuron devices in the in‐memory computing architecture, the advantages and challenges in area and energy efficiency are discussed by benchmarking various technologies. A small or even no membrane capacitor, a self‐reset property, and a high spiking frequency are highly desirable.

This study explores the design and control of the behaviour of agents and robots using simple circuits of spiking neurons and Spike Timing Dependent Plasticity (STDP) as a mechanism of associative and unsupervised learning. Based on a "reward and punishment" classical conditioning, it is demonstrated that these robots learnt to identify and avoid obstacles as well as to identify and look for rewarding stimuli. Using the simulation and programming environment NetLogo, a software engine for the Integrate and Fire model was developed, which allowed us to monitor in discrete time steps the dynamics of each single neuron, synapse and spike in the proposed neural networks. These spiking neural networks (SNN) served as simple brains for the experimental robots. The Lego Mindstorms robot kit was used for the embodiment of the simulated agents. In this paper the topological building blocks are presented as well as the neural parameters required to reproduce the experiments. This paper summarizes the resulting behaviour as well as the observed dynamics of the neural circuits. The Internet-link to the NetLogo code is included in the annex., Comment: Paper submitted to the BICT 2015 Conference in New York City, United States

There is a growing requirement in computational neuroscience for tools that permit collaborative model building, model sharing, combining existing models into a larger system (multi-scale model integration), and are able to simulate models using a variety of simulation engines and hardware platforms. Layered XML model specification formats solve many of these problems, however they are difficult to write and visualise without tools. Here we describe a new graphical software tool, SpineCreator, which facilitates the creation and visualisation of layered models of point spiking neurons or rate coded neurons without requiring the need for programming. We demonstrate the tool through the reproduction and visualisation of published models and show simulation results using code generation interfaced directly into SpineCreator. As a unique application for the graphical creation of neural networks, SpineCreator represents an important step forward for neuronal modelling.

The scientific interest attracted by Spiking Neural Networks (SNN) has lead to the development of tools for the simulation and study of neuronal dynamics ranging from phenomenological models to the more sophisticated and biologically accurate Hodgkin-and-Huxley-based and multi-compartmental models. However, despite the multiple features offered by neural modelling tools, their integration with environments for the simulation of robots and agents can be challenging and time consuming. The implementation of artificial neural circuits to control robots generally involves the following tasks: (1) understanding the simulation tools, (2) creating the neural circuit in the neural simulator, (3) linking the simulated neural circuit with the environment of the agent and (4) programming the appropriate interface in the robot or agent to use the neural controller. The accomplishment of the above-mentioned tasks can be challenging, especially for undergraduate students or novice researchers. This paper presents an alternative tool which facilitates the simulation of simple SNN circuits using the multi-agent simulation and the programming environment Netlogo (educational software that simplifies the study and experimentation of complex systems). The engine proposed and implemented in Netlogo for the simulation of a functional model of SNN is a simplification of integrate and fire (I&F) models. The characteristics of the engine (including neuronal dynamics, STDP learning and synaptic delay) are demonstrated through the implementation of an agent representing an artificial insect controlled by a simple neural circuit. The setup of the experiment and its outcomes are described in this work.

This paper studies dynamic of digital spike maps: simple digital dynamical systems that can generate various periodic spike-trains. In order to consider the periodic spike-trains, we presents two feature quantities: the number of co-existing periodic spike-trains late, the concentricity of periodic spike-trains. Using these two feature quantities. we construct the feature uantities plane. The feature quantities plane is useful in visualization, classification and consideration of the spike-train dynamics. The usefulness is confirmed in numerical experiment of typical example based on the analog bifurcating neurons.

© 2018 American Scientific Publishers. All rights reserved. Heartbeat classification using electrocardiogram (ECG) data is an essential feature of modern day wearable devices. State-of-the-art machine learning-based heartbeat classifiers are designed using convolutional neural networks (CNN). Despite their high classification accuracy, CNNs require significant computational resources and power. This makes the mapping of CNNs on resource- And power-constrained wearable devices challenging. In this paper, we propose heartbeat classification using spiking neural networks (SNN), an alternative approach based on a biologically inspired, event-driven neural networks. SNNs compute and transfer information using discrete spikes that require fewer operations and less complex hardware resources, making them energy-efficient compared to CNNs. However, due to complex error-backpropagation involving spikes, supervised learning of deep SNNs remains challenging. We propose an alternative approach to SNN-based heartbeat classification. We start with an optimized CNN implementation of the heartbeat classification task and then convert the CNN operations, such as multiply-accumulate, pooling and softmax, into spiking equivalent with a minimal loss of accuracy. We evaluate the SNN-based heartbeat classification using publicly available ECG database of the Massachusetts Institute of Technology and Beth Israel Hospital (MIT/BIH), and demonstrate a minimal loss in accuracy when compared to 85.92% accuracy of a CNN-based hearbeat classification. We demonstrate that, for every operation, the activation of SNN neurons in each layer is sparse when compared to CNN neurons, in the same layer. We also show that this sparsity increases with an increase in the number of layers of the neural network. In addition, we detail the power-accuracy trade-off of the SNN and show a 87.76% and 96.82% reduction in SNN neuron and synapse activity, respectively, for accuracy loss ranging between 0.6% and 1.00%, when compared to a CNN-only implementation. ispartof: JOURNAL OF LOW POWER ELECTRONICS vol:14 issue:4 pages:508-519 status: published

Neuronal network models and corresponding computer simulations are invaluable tools to aid the interpretation of the relationship between neuron properties, connectivity, and measured activity in cortical tissue. Spatiotemporal patterns of activity propagating across the cortical surface as observed experimentally can for example be described by neuronal network models with layered geometry and distance-dependent connectivity. In order to cover the surface area captured by today's experimental techniques and to achieve sufficient self-consistency, such models contain millions of nerve cells. The interpretation of the resulting stream of multi-modal and multi-dimensional simulation data calls for integrating interactive visualization steps into existing simulation-analysis workflows. Here, we present a set of interactive visualization concepts called views for the visual analysis of activity data in topological network models, and a corresponding reference implementation VIOLA (VIsualization Of Layer Activity). The software is a lightweight, open-source, web-based, and platform-independent application combining and adapting modern interactive visualization paradigms, such as coordinated multiple views, for massively parallel neurophysiological data. For a use-case demonstration we consider spiking activity data of a two-population, layered point-neuron network model incorporating distance-dependent connectivity subject to a spatially confined excitation originating from an external population. With the multiple coordinated views, an explorative and qualitative assessment of the spatiotemporal features of neuronal activity can be performed upfront of a detailed quantitative data analysis of specific aspects of the data. Interactive multi-view analysis therefore assists existing data analysis workflows. Furthermore, ongoing efforts including the European Human Brain Project aim at providing online user portals for integrated model development, simulation, analysis, and provenance tracking, wherein interactive visual analysis tools are one component. Browser-compatible, web-technology based solutions are therefore required. Within this scope, with VIOLA we provide a first prototype.

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.

Spiking neural networks are being investigated both as biologically plausible models of neural computation and also as a potentially more efficient type of neural network. Recurrent neural networks in the form of networks of gating memory cells have been central in state-of-the-art solutions in problem domains that involve sequence recognition or generation. Here, we design an analog Long Short-Term Memory (LSTM) cell where its neurons can be substituted with efficient spiking neurons, where we use subtractive gating (following the subLSTM in [1]) instead of multiplicative gating. Subtractive gating allows for a less sensitive gating mechanism, critical when using spiking neurons. By using fast adapting spiking neurons with a smoothed Rectified Linear Unit (ReLU)-like effective activation function, we show that then an accurate conversion from an analog subLSTM to a continuous-time spiking subLSTM is possible. This architecture results in memory networks that compute very efficiently, with low average firing rates comparable to those in biological neurons, while operating in continuous time.

Visual motion detection is essential for the survival of many species. The phenomenon includes several spatial properties, not fully understood at the level of a neural circuit. This paper proposes a computational model of a visual motion detector that integrates direction and orientation selectivity features. A recent experiment in the Drosophila model highlights that stimulus orientation influences the neural response of direction cells. However, this interaction and the significance at the behavioral level are currently unknown. As such, another objective of this article is to study the effect of merging these two visual processes when contextualized in a neuro-robotic model and an operant conditioning procedure. In this work, the learning task was solved using an artificial spiking neural network, acting as the brain controller for virtual and physical robots, showing a behavior modulation from the integration of both visual processes.

Cell assemblies are considered to be physiological as well as functional units in the brain. A repetitive and stereotypical sequential activation of many neurons was observed, but the mechanisms underlying it are not well understood. Feedforward networks, such as synfire chains, with the pools of excitatory neurons unidirectionally connected and facilitating signal transmission in a cascade-like fashion were proposed to model such sequential activity. When embedded in a recurrent network, these were shown to destabilise the whole network’s activity, challenging the suitability of the model. Here, we investigate a feedforward chain of excitatory pools enriched by inhibitory pools that provide disynaptic feedforward inhibition. We show that when embedded in a recurrent network of spiking neurons, such an augmented chain is capable of robust signal propagation. We then investigate the influence of overlapping two chains on the signal transmission as well as the stability of the host network. While shared excitatory pools turn out to be detrimental to global stability, inhibitory overlap implicitly realises the motif of lateral inhibition, which, if moderate, maintains the stability but if substantial, it silences the whole network activity including the signal. Addition of a disinhibitory pathway along the chain proves to rescue the signal transmission by transforming a strong inhibitory wave into a disinhibitory one, which specifically guards the excitatory pools from receiving excessive inhibition and thereby allowing them to remain responsive to the forthcoming activation. Disinhibitory circuits not only improve the signal transmission, but can also control it via a gating mechanism. We demonstrate that by manipulating a firing threshold of the disinhibitory neurons, the signal transmission can be enabled or completely blocked. This mechanism corresponds to cholinergic modulation, which was shown to be signalled by volume as well as phasic transmission and variably target classes of neurons. Furthermore, we show that modulation of the feedforward inhibition circuit can promote generating spontaneous replay at the absence of external inputs. This mechanism, however, tends to also cause global instabilities. Overall, these results underscore the importance of inhibitory neuron populations in controlling signal propagation in cell assemblies as well as global stability. Specific inhibitory circuits, when controlled by neuromodulatory systems, can robustly guide or block the signals and invoke replay. This mounts to evidence that the population of interneurons is diverse and can be best categorised by neurons’ specific circuit functions as well as their responsiveness to neuromodulators.

Simulation of brain neurons in real-time using biophysically meaningful models is a prerequisite for comprehensive understanding of how neurons process information and communicate with each other, in effect efficiently complementing in-vivo experiments. State-of-the-art neuron simulators are, however, capable of simulating at most few tens/hundreds of biophysically accurate neurons in real-time due to the exponential growth in the interneuron communication costs with the number of simulated neurons. In this paper, we propose a real-time, reconfigurable, multichip system architecture based on localized communication, which effectively reduces the communication cost to a linear growth. All parts of the system are generated automatically, based on the neuron connectivity scheme. Experimental results indicate that the proposed system architecture allows the capacity of over 3000 to 19 200 (depending on the connectivity scheme) biophysically accurate neurons over multiple chips.

How is information represented in real neural networks? Experimental results continue to provide evidence for presence of spiking patterns in network activity. The concept of polychronous groups attempts to explain these results by proposing that neurons group together to fire in non- synchronous but precise time-locked chains. Several methods for the detection of such groups have been proposed, however, they all employ extensive searching in network structure, which limits their usefulness. We present a new method by observing spiking dependencies in network activity to directly detect polychronous groups. Our method shows comparatively more efficient computation by trading off detection selectivity. The method allows for analysis of polychronous groups emerging in noisy networks. Our results support the existence of structure-forming properties of spontaneous activity in neural network.

The massively parallel nature of biological information processing plays an important role for its superiority to human-engineered computing devices. In particular, it may hold the key to overcoming the von Neumann bottleneck that limits contemporary computer architectures. Physical-model neuromorphic devices seek to replicate not only this inherent parallelism, but also aspects of its microscopic dynamics in analog circuits emulating neurons and synapses. However, these machines require network models that are not only adept at solving particular tasks, but that can also cope with the inherent imperfections of analog substrates. We present a spiking network model that performs Bayesian inference through sampling on the BrainScaleS neuromorphic platform, where we use it for generative and discriminative computations on visual data. By illustrating its functionality on this platform, we implicitly demonstrate its robustness to various substrate-specific distortive effects, as well as its accelerated capability for computation. These results showcase the advantages of brain-inspired physical computation and provide important building blocks for large-scale neuromorphic applications., Comment: This preprint has been published 2019 November 14. Please cite as: Kungl A. F. et al. (2019) Accelerated Physical Emulation of Bayesian Inference in Spiking Neural Networks. Front. Neurosci. 13:1201. doi: 10.3389/fnins.2019.01201

Els models neuronals de camp mig són descripcions fenomenològiques de l'activitat de xarxes de neurones espacialment organitzades. Gràcies a la seva simplicitat, aquests models són unes eines extremadament útils per a l'anàlisi dels patrons espai-temporals que apareixen a les xarxes neuronals, i s'utilitzen àmpliament en neurociència computacional. És ben sabut que els models de camp mig tradicionals no descriuen adequadament la dinàmica de les xarxes de neurones si aquestes actuen de manera síncrona. No obstant això, les simulacions computacionals de xarxes neuronals demostren que, fins i tot en estats d'alta asincronia, fluctuacions ràpides dels inputs comuns que arriben a les neurones poden provocar períodes transitoris en els quals les neurones de la xarxa es comporten de manera síncrona. A més a més, la sincronització també pot ser generada per la mateixa xarxa, donant lloc a oscil·lacions auto-sostingudes. En aquesta tesi investiguem la presència de patrons espai-temporals deguts a la sincronització en xarxes de neurones heterogènies i espacialment distribuïdes. Aquests patrons no s'observen en els models tradicionals de camp mig, i per aquest motiu han estat àmpliament ignorats en la literatura. Per poder investigar la dinàmica induïda per l'activitat sincronitzada de les neurones, fem servir un nou model de camp mig que es deriva exactament d'una població de neurones de tipus quadratic integrate-and-fire. La simplicitat del model ens permet analitzar l'estabilitat de la xarxa en termes del perfil espacial de la connectivitat sinàptica, i obtenir fórmules exactes per les fronteres d'estabilitat que caracteritzen la dinàmica de la xarxa neuronal original. Aquest mateix anàlisi també revela l'existència d'un conjunt de modes d'oscil·lació que es deuen exclusivament a l'activitat sincronitzada de les neurones. Creiem que els resultats presentats en aquesta tesi inspiraran nous avenços teòrics relacionats amb la dinàmica col·lectiva de les xarxes neuronals, contribuïnt així en el desenvolupament de la neurociència computacional. Neural field models are phenomenological descriptions of the activity of spatially organized, recurrently coupled neuronal networks. Due to their mathematical simplicity, such models are extremely useful for the analysis of spatiotemporal phenomena in networks of spiking neurons, and are largely used in computational neuroscience. Nevertheless, it is well known that traditional neural field descriptions fail to describe the collective dynamics of networks of synchronously spiking neurons. Yet, numerical simulations of networks of spiking neurons show that, even in the case of highly asynchronous activity, fast fluctuations in the common external inputs drive transient episodes of spike synchrony. Moreover, synchronization may also be generated by the network itself, resulting in the appearance of robust large-scale, self-sustained oscillations. In this thesis, we investigate the emergence of synchrony-induced spatiotemporal patterns in spatially distributed networks of heterogeneous spiking neurons. These patterns are not observed in traditional neural field theories and have been largely overlooked in the literature. To investigate synchrony-induced phenomena in neuronal networks, we use a novel neural field model which is exactly derived from a large population of quadratic integrate-and-fire model neurons. The simplicity of the neural field model allows us to analyze the stability of the network in terms of the spatial profile of the synaptic connectivity, and to obtain exact formulas for the stability boundaries characterizing the dynamics of the original spiking neuronal network. Remarkably, the analysis also reveals the existence of a collection of oscillation modes, which are exclusively due to spike-synchronization. We believe that the results presented in this thesis will foster theoretical advances on the collective dynamics of neuronal networks, upgrading the mathematical basis of computational neuroscience.

Neuronal network models and corresponding computer simulations are invaluable tools to aid the interpretation of the relationship between neuron properties, connectivity and measured activity in cortical tissue. Spatiotemporal patterns of activity propagating across the cortical surface as observed experimentally can for example be described by neuronal network models with layered geometry and distance-dependent connectivity. The interpretation of the resulting stream of multi-modal and multi-dimensional simulation data calls for integrating interactive visualization steps into existing simulation-analysis workflows. Here, we present a set of interactive visualization concepts called views for the visual analysis of activity data in topological network models, and a corresponding reference implementation VIOLA (VIsualization Of Layer Activity). The software is a lightweight, open-source, web-based and platform-independent application combining and adapting modern interactive visualization paradigms, such as coordinated multiple views, for massively parallel neurophysiological data. For a use-case demonstration we consider spiking activity data of a two-population, layered point-neuron network model subject to a spatially confined excitation originating from an external population. With the multiple coordinated views, an explorative and qualitative assessment of the spatiotemporal features of neuronal activity can be performed upfront of a detailed quantitative data analysis of specific aspects of the data. Furthermore, ongoing efforts including the European Human Brain Project aim at providing online user portals for integrated model development, simulation, analysis and provenance tracking, wherein interactive visual analysis tools are one component. Browser-compatible, web-technology based solutions are therefore required. Within this scope, with VIOLA we provide a first prototype., Comment: 38 pages, 10 figures, 3 tables

Frontiers in neuroinformatics 12, 68 (2018). doi:10.3389/fninf.2018.00068, Published by Frontiers Research Foundation, Lausanne

As the brain purportedly employs on-policy reinforcement learning compatible with SARSA learning, and most interesting cognitive tasks require some form of memory while taking place in continuous-time, recent work has developed plausible reinforcement learning schemes that are compatible with these requirements. Lacking is a formulation of both computation and learning in terms of spiking neurons. Such a formulation creates both a closer mapping to biology, and also expresses such learning in terms of asynchronous and sparse neural computation. We present a spiking neural network with memory that learns cognitive tasks in continuous time. Learning is biologically plausibly implemented using the AuGMeNT framework, and we show how separate spiking forward and feedback networks suffice for learning the tasks just as fast the analog CT-AuGMeNT counterpart, while computing efficiently using very few spikes: 1–20 Hz on average.

Learning and reproducing temporal sequences is a fundamental ability used by living beings to adapt behaviour repertoire to environmental constraints. This paper is focused on the description of a model based on spiking neurons, able to learn and autonomously generate a sequence of events. The neural architecture is inspired by the insect Mushroom Bodies (MBs) that are a crucial centre for multimodal sensory integration and behaviour modulation. The sequence learning capability coexists, within the insect brain computational model, with all the other features already addressed like attention, expectation, learning classification and others. This is a clear example that a unique neural structure is able to cope concurrently with a plethora of behaviours. Simulation results and robotic experiments are reported and discussed.

Constraint satisfaction problems (CSP) are at the core of numerous scientific and technological applications. However, CSPs belong to the NP-complete complexity class, for which the existence (or not) of efficient algorithms remains a major unsolved question in computational complexity theory. In the face of this fundamental difficulty heuristics and approximation methods are used to approach instances of NP (e.g., decision and hard optimization problems). The human brain efficiently handles CSPs both in perception and behavior using spiking neural networks (SNNs), and recent studies have demonstrated that the noise embedded within an SNN can be used as a computational resource to solve CSPs. Here, we provide a software framework for the implementation of such noisy neural solvers on the SpiNNaker massively parallel neuromorphic hardware, further demonstrating their potential to implement a stochastic search that solves instances of P and NP problems expressed as CSPs. This facilitates the exploration of new optimization strategies and the understanding of the computational abilities of SNNs. We demonstrate the basic principles of the framework by solving difficult instances of the Sudoku puzzle and of the map color problem, and explore its application to spin glasses. The solver works as a stochastic dynamical system, which is attracted by the configuration that solves the CSP. The noise allows an optimal exploration of the space of configurations, looking for the satisfiability of all the constraints; if applied discontinuously, it can also force the system to leap to a new random configuration effectively causing a restart.

Neuro-biology inspired Spiking Neural Network (SNN) enables efficient learning and recognition tasks. To achieve a large scale network akin to biology, a power and area efficient electronic neuron is essential. Earlier, we had demonstrated an LIF neuron by a novel 4-terminal impact ionization based n+/p/n+ with an extended gate (gated-INPN) device by physics simulation. Excellent improvement in area and power compared to conventional analog circuit implementations was observed. In this paper, we propose and experimentally demonstrate a compact conventional 3-terminal partially depleted (PD) SOI- MOSFET (100 nm gate length) to replace the 4-terminal gated-INPN device. Impact ionization (II) induced floating body effect in SOI-MOSFET is used to capture LIF neuron behavior to demonstrate spiking frequency dependence on input. MHz operation enables attractive hardware acceleration compared to biology. Overall, conventional PD-SOI-CMOS technology enables very-large-scale-integration (VLSI) which is essential for biology scale (~1011 neuron based) large neural networks.

The Self-Organizing Map (SOM) is a recurrent neural network topology that realizes competitive learning for the unsupervised classification of data. In this paper, we investigate the design of a spiking neural network-based SOM for the classification of bioelectric-timescale signals. We present novel insights into the architectural design space, inherent trade-offs, and the critical requirements for designing and configuring neurons, synapses and learning rules to achieve stable and accurate behaviour. We perform this exploration using high-level architectural simulations and, additionally, through the full-custom implementation of components.

International audience; Human working memory is capable to generate dynamically robust and flexible neuronal sequences for action planning, problem solving and decision making. However, current neurocomputational models of working memory find hard to achieve these capabilities since intrinsic noise is difficult to stabilize over time and destroys global synchrony. As part of the principle of free-energy minimization proposed by Karl Fris-ton, we propose a novel neural architecture to optimize the free-energy inherent to spiking recurrent neural networks to regulate their activity. We show for the first time that it is possible to stabilize iteratively the long-range control of a recurrent spiking neurons network over long sequences. We identify our architecture as the working memory composed by the Basal Ganglia and the Intra-Parietal Lobe for action selection and we make some comparisons with other networks such as deep neural networks and neural Turing machines. We name our architecture INFERNO for Iterative Free-Energy Optimization for Recurrent Neural Network.

Development of the state-of-the-art cyber-physical systems (CPS), which incorporate physical and computational components, poses new challenges for research and industry. In order for CPS serve its purpose to make human lives safer, easier, more enjoyable and convenient, both academia and industry needs to develop new methods for control and monitoring of such systems. Neural models are a prospective direction for design of CPS controllers and monitors, and in the thesis we first show, how neural models can be applied in CPS control to quantify the uncertainty of the system. We then show, how digital spiking neural model, called TrueNorth, can be used for runtime monitoring of temporal logic specifications of mission-critical properties. To be able to deliver not only qualitative verdict, but also reason quantitatively, we propose an approach to use the model for computation of arithmetic functions and implement neural monitors for semantics of temporal logic based on circular convolution. In the applied part of the work we show how runtime monitoring can speed up verification and validation phase in automotive electronic development. We identify phases, where runtime monitoring can facilitate both concept and post-silicon verification and testing. To build runtime monitors that are capable to keep up with the speed of the physical sensor, we developed an approach to convert formalized requirements to hardware monitors, which are then synthesized in FPGA. Reuse of the monitors from concept to post-silicon verification phase using high-level synthesis speeds up the testing process and enables long-term requirements evaluation. We illustrated our approach by formalizing, creating hardware monitors and evaluating the results in the lab environment for electrical and timing requirements of industrial SENT and SPC protocols.

In this paper, we present the numerical results of the implementation of a Spiking Central Pattern Generator (SCPG) on a SpiNNaker board. The SCPG is a network of current-based leaky integrateand- fire (LIF) neurons, which generates periodic spike trains that correspond to different locomotion gaits (i.e. walk, trot, run). To generate such patterns, the SCPG has been configured with different topologies, and its parameters have been experimentally estimated. To validate our designs, we have implemented them on the SpiNNaker board using PyNN and we have embedded it on a hexapod robot. The system includes a Dynamic Vision Sensor system able to command a pattern to the robot depending on the frequency of the events fired. The more activity the DVS produces, the faster that the pattern that is commanded will be. Ministerio de Economía y Competitividad TEC2016-77785-P