Search

With the rapid growth of AI and large language models, concerns about their energy consumption have become increasingly prominent. The traditional digital von Neumann architecture is ill-suited for deep neural networks, prompting the exploration of neuromorphic computational models. These models draw inspiration from nature and the brain to integrate memory and computation efficiently. This thesis, part of the EU-funded InsectNeuroNano project, aims to advance neuromorphic computing. DASA molecules, a type of photoswitching dye, are used in this project to replicate synaptic weights between neurons. This enables light to act as a signal carrier, allowing the dye to calculate inter-neuron weights. The project investigates dye deposition on 1 cm × 1 cm sapphire slabs, measuring changes in transmission over time and fitting a kinetic decay model. Additionally, this thesis explores the replication of biological neurons' abilities to receive excitatory and inhibitory signals, process them via a sigmoid activation function, and transmit outputs. As part of the InsectNeuroNano project, III-V nanowires have been combined to create a simple artificial neuron replicating these characteristics. Various measurements on these devices aim to demonstrate their ability to mimic the necessary properties of biological neurons. These measurements include electrical and photoelectrical aspects, divided into standardized and general tests. The nanowire neuron circuits demonstrated vital properties such as inhibition, excitation, sigmoid activation, and pulse trains. However, these effects were not fully optimized, indicating a gap between biological and artificial neurons. Additional unexplained effects were also observed. Only two samples fit the kinetic model for the dye experiments, with the other samples affected by irreversible bleaching and where the measurement continued to bleach them. The limited data prevented definitive conclusions about the deposition parameter space, though, Det är sommar och fåglarna flyger kvickt genom luften, lekande i de varma vindbyarna. För insekterna är detta som vanligt, en kamp mot en fientlig värld och oräkneliga varelser som vill äta dem. Myrorna skyndar sig fram över varma stenar och mellan vegetation i jakt på något som kan tas tillbaka som mat. Varje myra är förbrukningsbar, enbart kolonin spelar roll. Men trots den snålheten med energi som detta leder till klarar de inte bara att hitta mat utan också att leta sig tillbaka till kolonin. Deras små hjärnor, arbetande med den energin som de har, lyckas navigera genom denna ogästvänliga och stora värld. Gömd från sommarens hetta i den svalkande skuggan av en ek iakttar du myrorna. Men så väck en tanke. Är detta verkligen så imponerande? Har inte människan AI som med alla dess funktionaliteter och förmågor är vida överlägsna i intelligens hela myrornas kolonier? Svaret är inte riktigt så enkelt. AI klarar av att lösa många problem som är omöjliga för myrorna men har samtidigt stora begränsningar. AI, i all sin strålande ungdom, kräver mycket mer energi än insekter, något som blivit ett allt mer uppmärksammat problem. I förhållande har insekterna haft flera hundra miljoner år av evolution på sig att optimera sin intelligens. I vad många kallar ”age of AI” så har detta problem blivit allt mer framträdande och EU har genom olika projekt försökt finna lösningar. Denna uppsats är genomförd som en del av ett av dessa projekt, InsectNeuroNano, där målet är att skapa en fysisk kopia av insekternas navigationssystem. För att detta skall fungera krävs en bra förståelse över hur de planerade komponenterna fungerar vilket är där denna uppsats kommer in med sitt bidrag. Genom forskning på insekter vid Lunds universitet har forskare kartlagt en region i deras hjärna kallad det centrala komplexet, som är ansvarigt för navigation. För att kunna efterlikna denna funktion har forskarna använt nanotrådar. Dessa är små pelare som mäter ungefär en hundradel av ett hårstrå i längd o

Biological nervous systems constitute important sources of inspiration towards computers that are faster, cheaper, and more energy efficient. Neuromorphic disciplines view the brain as a coevolved system, simultaneously optimizing the hardware and the algorithms running on it. There are clear efficiency gains when bringing the computations into a physical substrate, but we presently lack theories to guide efficient implementations. Here, we present a principled computational model for neuromorphic systems in terms of spatio-temporal receptive fields, based on affine Gaussian kernels over space and leaky-integrator and leaky integrate-and-fire models over time. Our theory is provably covariant to spatial affine and temporal scaling transformations, and with close similarities to the visual processing in mammalian brains. We use these spatio-temporal receptive fields as a prior in an event-based vision task, and show that this improves the training of spiking networks, which otherwise is known as problematic for event-based vision. This work combines efforts within scale-space theory and computational neuroscience to identify theoretically well-founded ways to process spatio-temporal signals in neuromorphic systems. Our contributions are immediately relevant for signal processing and event-based vision, and can be extended to other processing tasks over space and time, such as memory and control., QC 20240508

Artificial Intelligence (AI) promises economic growth and solutions to global problems but also raises societal concerns. Training AI models has a big carbon footprint due to data processing in fossil-fuel-reliant data centers. If the data centers are outside the European legal space, data processing incurs privacy risks. Besides, reliance on AI aggravates Europe’s dependence on non-European chip-makers, whose supply chains can be disrupted. To address such concerns, NeuroSys develops energy-efficient neuromorphic hardware tailored to AI applications that protect privacy by processing data locally. NeuroSys aims to build a chip plant near Aachen in Germany to support Europe’s technological sovereignty. This depends on an innovation ecosystem where socio-technical transformations emerge in transdisciplinary collaboration. This chapter introduces NeuroSys as a testbed for studying how transformation research can contribute to the sustainability and trustworthiness of AI Made in Europe.

In recent years, the rapid advancement of Artificial Intelligence (AI) technologies has produced a significant impact on our society. However, as AI systems increase in complexity, they encounter significant challenges, especially concerning energy efficiency and computational demands. These challenges predominantly stem from the reliance of AI systems on conventional digital processors, which are based on the von Neumann architecture. To overcome these limitations, there is a growing interest in exploring unconventional computing methods that leverage diverse physical substrates for computation. A possible source of inspiration comes from biological neural networks, which are far more efficient than their artificial counterparts. In this context, Reservoir Computing has emerged as one of the bio-inspired alternatives to implement machine learning using physical substrates. This PhD thesis centers on benchmarking Reservoir Computing across new applications and exploring innovative methodologies. At the heart of my research is the experimental implementation of an optoelectronic reservoir computer, which combines photonic hardware with a high-speed electronic device, a field-programmable gate array board. This experimental system is used to address three research questions. (1) The development of a training procedure called ‘Timesteps of Interest’ to solve a complex task, the recognition of human actions in video sequences. This approach achieves results that are comparable to those obtained with more complex neural networks implemented on conventional digital hardware. (2) The investigation of Reservoir Computing architectures featuring hierarchical interconnections of reservoirs, which demonstrate enhanced performance over standard Reservoir Computing schemes. (3) The adoption of a new optimization technique for the reservoir, based on the use of a delayed input. This method proves to be simpler and more effective than the optimization traditionally used for Reserv, Ces dernières années, le développement rapide des technologies de l'Intelligence Artificielle (IA) a eu un impact significatif sur notre société. Cependant, à mesure que les systèmes d'IA deviennent plus complexes, ils rencontrent d'importants défis, notamment en termes d'efficacité énergétique et de demandes computationnelles. Ces défis découlent principalement du fait que les systèmes d'IA dépendent des processeurs numériques conventionnels, qui sont basés sur l'architecture de von Neumann. Pour surmonter ces limitations, un intérêt croissant est porté à l'exploration de méthodes de calcul non conventionnelles qui tirent parti de substrats physiques divers pour le calcul. Une source possible d'inspiration provient des réseaux neuronaux biologiques, qui sont bien plus efficaces que leurs homologues artificiels. Dans ce contexte, le Reservoir Computing est l'une des alternatives bio-inspirées actuellement étudiée pour mettre en œuvre l'apprentissage automatique à l'aide de substrats physiques.Cette thèse de doctorat est axée sur l’étude de nouvelles applications du reservoir computing et sur l'exploration de méthodologies innovantes. Au cœur de ma recherche se trouve la mise en œuvre expérimentale d'un reservoir computer optoélectronique, qui combine du matériel photonique avec un dispositif électronique à grande vitesse, un field-programmable gate array. Ce système expérimental est utilisé pour répondre à trois questions de recherche. (1) Le développement d'une procédure d'entraînement appelée "Timesteps of Interest" pour résoudre une tâche complexe, la reconnaissance d'actions humaines dans des séquences vidéo. Cette approche permet d'obtenir des résultats comparables à ceux obtenus avec des réseaux neuronaux plus complexes mis en œuvre sur du matériel numérique conventionnel. (2) L'étude d'architectures de reservoir computing comportant des interconnexions hiérarchiques de réservoirs, qui démontrent des performances accrues par rapport aux schémas standard. (3) L, Doctorat en Sciences, E.P acknowledges financial support from the H2020 Marie Skłodowska-Curie Actions (Project POSTDIGITAL Grant number 860360), and from the Fonds de la Recherche Scientifique - FNRS, Belgium (funding EOS n° 40007536 and CDR n° 40021243)., info:eu-repo/semantics/nonPublished

This thesis lies within the field of unconventional computing, and in particular focuses on physical systems whose dynamics can usefully process information.Such platforms may not outperform an electronic processor but may offer lower power consumption and better integration with sensors, and consequently may be useful for edge computing applications.Here, we study photonic systems, leveraging the relative maturity of the field due to the growth of telecommunication-related industries.Specifically, we explore how the wavelength degree of freedom of the light can be exploited to implement frequency multiplexed Artificial Neural Network (ANN) algorithms.Among many kinds of ANN algorithms, we focus on randomized ANNs. These networks are mostly composed of randomly selected and untrained connections, and only the output connections are tuned. Randomized ANNs offer interesting advantages, such as the possibility of low-cost training (both in terms of time and power) and the simplicity of implementation in physical computers. We describe the implementation of three randomized ANN algorithms: Extreme Learning Machine (ELM), Reservoir Computer (RC), and Deep Reservoir Computer (Deep-RC).We describe a variety of schemes where artificial neurons are encoded in the lines of an optical frequency comb.The mixing of information, necessary for machine learning, happens either using periodic phase modulation or using Kerr nonlinearity.Output weights can be applied optically by using programmable spectral filters.Each scheme is experimentally demonstrated in a fiber-based setup and evaluated on standard machine learning benchmark tasks.Benchmark results are in line with the performance reported on similar photonic-based systems and equivalent software implementations. We demonstrate that parallel operations can be executed on a single hardware.Parallel-running computations can be concatenated, thus constituting the layers of a Deep-RC.To the best of our knowledge, we report the firs, Cette thèse s'inscrit dans le domaine de l'informatique non conventionnelle et se concentre sur les systèmes physiques dont la dynamique peut être utilisée pour traiter l'information.Ces plateformes ne sont pas forcément plus performantes qu'un processeur électronique, mais peuvent offrir une consommation d'énergie plus faible et une meilleure intégration avec les capteurs, et peuvent donc être utiles pour les applications de edge computing.Nous étudions ici des systèmes photoniques, en tirant parti de la maturité relative du domaine due à la croissance de l’industries des télécommunications.Plus précisément, nous étudions comment le degré de liberté de la longueur d'onde de la lumière peut être exploité pour mettre en œuvre des algorithmes de réseaux neuronaux artificiels multiplexés en fréquence.Parmi les nombreux types d'algorithmes de réseaux neuronaux artificiels, nous nous concentrons sur les architectures randomisées.Ces réseaux sont principalement composés de connexions sélectionnées au hasard et non entraînées, et seules les connexions de sortie sont ajustées.Ces architectures randomisées offrent des avantages intéressants, tels que la possibilité d'un entraînement peu coûteux (à la fois en termes de temps et d'énergie) et la simplicité de mise en œuvre dans les ordinateurs physiques.Nous décrivons la mise en œuvre de trois algorithmes :l’Extreme Learning Machine (ELM), le Reservoir Computer (RC) et le Reservoir Computer Profond (Deep-RC).Nous décrivons une variété de schémas dans lesquels les neurones artificiels sont encodés dans les lignes d'un peigne de fréquence optique.Le mélange des informations, nécessaire à l'apprentissage automatique, se fait soit par modulation de phase périodique, soit par non-linéarité de Kerr.Les poids de sortie peuvent être appliqués optiquement à l'aide de filtres spectraux programmables. Chaque schéma fait l'objet d'une démonstration expérimentale en fibre optique et est évalué sur des tâches de référence standard en matièr, Doctorat en Sciences, info:eu-repo/semantics/nonPublished

Multimodal perception can capture more precise and comprehensive information compared with unimodal approaches. However, current sensory systems typically merge multimodal signals at computing terminals following parallel processing and transmission, which results in the potential loss of spatial association information and requires time stamps to maintain temporal coherence for time-series data. Here we demonstrate bioinspired in-sensor multimodal fusion, which effectively enhances comprehensive perception and reduces the level of data transfer between sensory terminal and computation units. By adopting floating gate phototransistors with reconfigurable photoresponse plasticity, we realize the agile spatial and spatiotemporal fusion under nonvolatile and volatile photoresponse modes. To realize an optimal spatial estimation, we integrate spatial information from visual-tactile signals. For dynamic events, we capture and fuse in real time spatiotemporal information from visual-audio signals, realizing a dance-music synchronization recognition task without a time-stamping process. This in-sensor multimodal fusion approach provides the potential to simplify the multimodal integration system, extending the in-sensor computing paradigm. © 2024 American Chemical Society.

Despite its successes, to date Artificial Intelligence (AI) is still characterized by a number of shortcomings with regards to different application domains and goals. These limitations are arguably both conceptual (e.g., related to the underlying theoretical models, such as symbolic vs.connectionist), and operational (e.g., related to robustness and ability to generalize). Biologically inspired AI, and more specifically brain-inspired AI, promises to provide further biological aspects beyond those that are already traditionally included in AI, making it possible to assess and possibly overcome some of its present shortcomings. This article examines some conceptual, technical, and ethical issues raised by the development and use of brain-inspired AI. Against this background, the paper asks whether there is anything ethically unique about brain-inspired AI. The aim of the paper is to introduce a method that has a heuristic nature and that can be applied to identify and address the ethical issues arising from brain-inspired AI (and from AI more generally). The conclusion resulting from the application of this method is that, compared to traditional AI, brain-inspired AI raises new foundational ethical issues and some new practical ethical issues, and exacerbates some of the issues raised by traditional AI.

Near-Infrared (NIR) Spectroscopy is widely applied in agriculture and food industry for the determination of fruit ripeness, the content in soluble solids, pH and acidity. In this study, we report on the develeopment of a novel neuromorphic classifier based on Spiking Neural Networks (SNNs) to classify NIR spectra of fruit species. Neuromorphic computing holds the potential for a low-power real-time recognition system based on NIR spectroscopy signals that could be used not only in food industry, but also in pharmaceutical and medical applications. For benchmarking, we compare the performance of the classifier to the performance of non-spiking Artificial Neural Networks (ANN) and Support Vector Machines (SVM). Furthermore, we show how our SNN-based algorithm can be implemented in the mixedsignal analog-digital neuromorphic device DYNAP-SE.

Neuromorphic computational networks, inspired by biological neural networks, provide a possible way of lowering computational energy cost, while at the same time allowing for much more sophisticated devices capable of real-time inferences and learning. Since simulating artificial neural networks on conventional computers is particularly inefficient, the development of neuromorphic devices is strongly motivated as the reliance on AI-models increases. A neuromorphic device inspired by the insect brain and capable of navigation, has previously been proposed, where optically communicating semiconductor nanowires form the basis of the network, and a state-of-the-art molecular dye exhibiting reversible bleaching provides the synaptic weights. This Thesis explores integral sub-components of that device, contributing concrete results towards its realization. The molecular dye is combined with a nanowire array solar cell device in order to fashion a high-sensitivity nanoscale device with integrated memory functionality. Compared to purely electronic "memory-resistors", this approach features a greater potential for multiplexing due to the wavelength sensitivity of different molecular dye species. The effects of dye concentration, deposition method, film thickness and annealing on the molecular photophysics are explored, identifying starting points for a design optimization process. A high degree of repeatability of the memory functionality is shown, even at prolonged high-intensity illumination. Additionally, opto-electronic single nanowire ciruits, where the open-circuit voltage generated by the illumination of a p-i-n InP nanowire gates an InAs nanowire, are characterized in high-precision Optical Beam Induced Current experiments. Depending on the orientation of the InP nanowire, increase or decrease of the InAs conductance is achieved, suggesting that the circuits could be implemented as excitatory and inhibitory nodes in a nanowire-based neural network. Finally, the opti, Det är en varm och solig dag. Du ligger på rygg på en grönskande äng och ur ögonvrån ser du ett bi surra av och an tätt ovanför marken, på jakt efter nektar och pollen. Biet tar ett varv runt ängen medan du ligger kvar och funderar ett slag över hur ett sådant bi, som måste flyga hit och dit för att leta upp blommande växter, förmår att hitta hem igen. Dina tankar avbryts av ett surrande: biet närmar sig igen. Men den här gången flyger det målmedvetet fram över ängen, inte längre dansande fram utan i en nästan rät linje rakt ovanför ditt huvud. Om du hade gissat att biet nu var på väg hem till sin kupa, så hade du gissat rätt. Det är nämligen så att bolevande insekter så som bin, humlor och myror klarar av att ta sig raka vägen hem, oavsett hur krokig vägen under deras utfärd varit. På något sätt måste de hela tiden kunna hålla koll på åt vilket håll "hem" är. Svaret på frågan om hur ett bi förmår hitta hem är komplicerat, men forskare i Lund har lyckats kartlägga delar av humlors hjärna, och där identifierat det nätverk av nervceller som tros ansvara för just denna typ av navigation. Forskningen som presenteras i denna uppsats syftar till att praktiskt undersöka om man skulle kunna bygga en kopia av detta särskilda nätverk med hjälp av byggstenar som inte är lika fullt komplicerade som de nervceller som utgör insekthjärnan. En av dessa byggstenar är så kallade nanotrådar: en typ av små pinnar mycket tunnare än ett hårstrå. Dessa "nanopinnar" kan fungera både som solceller och LED-lampor. Om man är klyftig kan man placera ut dem på ett sådant vis att varje enskild nanotråd känner av ljuset som de andra skickar ut, i större eller mindre utsträckning. Detta liknar på ett sätt kommunikationen som sker mellan nervceller i en hjärna. Men insekthjärnan, precis som vår egen hjärna, har ett viktigt knep till sitt förfogande: det är nämligen så att kopplingarna mellan nervceller som ofta pratar med varandra stärks, och tvärtom. På det viset kan biet komma ihåg varåt hemmet l

The backbone of modern computing systems rely on two key things: logic and memory, and while computing power hasseen tremendous advancements through scaling of the fundamental building block – the transistor, memory access hasn’tevolved as rapidly, leading to significant memory-bound systems. Additionally, the rapid evolution of machine learningand deep neural network (DNN) applications, has exposed the fundamental limitations of the traditional von Neumanncomputing architecture, due to its heavy reliance on memory access. The physical separation between the computing unitand the memory in von Neumann architectures is limiting performance and energy efficiency. A promising solution toaddress these challenges is the development of emerging non-volatile memory technologies that provide significant scalingand integration possibilities, fast switching speeds, and highly energy-efficient operations. Additionally, by integrating“memory resistors” (memristors) in large crossbar arrays, the computation can take place in-memory which can resolve thebottleneck in traditional von Neumann architectures.This thesis investigates the implementation of ferroelectric HfO2 in ferroelectric tunnel junctions (FTJs) and ferroelectricfield effect transistors (FeFETs) as potential candidates for emerging non-volatile memories and memristors.Initially, the thesis focuses on the integration of ferroelectric HfO2 onto the high mobility III-V semiconductor InAs forthe fabrication of metal-oxide-semiconductor (MOS) capacitors. Moreover, optimization of the processing conditions on thecritical interface between the semiconductor and high-k oxide is extensively studied using both electrical characterization andsynchrotron radiation techniques. After optimization of the annealing treatment and top electrode texturing, the fabricationof vertical InAs nanowire FeFETs is successfully implemented. The FeFET shows encouraging initial results with limitationssolvable by further process engineering.The

Neuromorphic computing can be used to efficiently implement spiking neural networks. Such spiking neural networks can be used in edge AI applications, where low power consumption is paramount. The use of analog components allows for extremely low power implementations. This thesis contributes the designs of an analog spike generator, synaptic elements and an accumulating neuron in 28 nm CMOS technology. The elements are assembled in a neural network and laid out in an SoC. Energy consumption numbers of less than 1 pJ/synaptic operation are achieved in the analog neuromorphic components., Electrical Engineering | Circuits and Systems

Autonomous vehicles promise to revolutionize society and improve the daily life of many, making them a coveted aim for a vast research community. To enable complex reasoning in autonomous vehicles, researchers are exploring new methods beyond traditional engineering approaches, in particular the idea of drawing inspiration from the only existing being able to drive: the human. The mental processes behind the human ability to drive can inspire new approaches with the potential to bridge the gap between artificial drivers and human drivers. In this review, we categorize and evaluate existing work on autonomous driving influenced by cognitive science, neuroscience, and psychology. We propose a taxonomy of the various sources of inspiration and identify the potential advantages with respect to traditional approaches. Although these human-inspired methods have not yet reached widespread adoption, we believe they are critical to the future of fully autonomous vehicles., CC BY-NC-ND 4.0 DEED© 2023Corresponding author: E-mail address: alice.plebe@unitn.it (A. Plebe)

Enabled by the availability of both excitatory and inhibitory post-synaptic currents, a biological neural network is inherently capable of implementing more sophisticated non-monotonic classification schemes. While such currents are readily emulated in a software-based artificial neural network using both positive and negative synaptic weighting factors, the same is not as straight forward in a hardware implementation. In this work, two dual-gate thin-film transistors with effectively infinite direct-current input impedance are deployed to construct the excitatory and inhibitory conduction channels of an artificial synapse. The utility of a hardware-based artificial neural network constructed using such synapses is demonstrated by its deployment in the implementation of the complete set of sixteen 2-input binary logic functions exhibiting both monotonic and non-monotonic behavior. The set of weighting factors needed for the implementation of each function are determined using a neuromorphic feed-forward training algorithm based on gradient-descent. While some of the functions can be implemented using the simplest reconfigurable 2×1 network, all 16 of the functions can be implemented using a deeper, reconfigurable 3×2×1 network.

System designers are continuously exploring design methodologies that harness increased levels of heterogeneity towards pushing the boundaries of achievable performance gains. Optimizing a heterogenous computing system is a multi-dimensional and complex design space exploration problem particularly in scenarios where algorithms, data types, and processing resources and their availability vary dynamically. This dissertation addresses the need for productive application development and deployment on heterogenous systems through two open-source ecosystems with an aim to make emerging computing systems accessible for users. We have developed CEDR, an open-source, unified compilation, and runtime framework designed for heterogeneous systems. We introduce a new approach to runtime system development that reduces the overhead dedicated to managing system resources from 2000 cycles scale to 30 cycles scale, making CEDR practical for deployment in real-time embedded systems. This framework empowers users to develop, compile, and deploy applications on off-the-shelf heterogeneous computing platforms seamlessly, eliminating the need for users to possess specialized hardware expertise throughout the process. Furthermore, it facilitates the evaluation of resource management strategies and validation of hardware configurations within a single, cohesive framework. Importantly, this framework is portable across a wide range of Linux-based systems, ensuring that effort to migrate across systems is minimal for all developers involved. Its utility has been successfully tested by numerous partners from both industry and academia. With the same motivation we have developed RANC, an open-source highly flexible ecosystem that enables rapid experimentation with neuromorphic architectures in both software via C++ simulation and hardware via FPGA emulation. RANC allows optimizing architectures based on application insights as well as prototyping future neuromorphic architectures that can sup

Recently, Spiking Neural Networks (SNNs) have been proposed as alternatives to Deep Neural Networks (DNNs) for low-power processing of sensory data. SNNs use neurons which do not output real-valued activations, but a sequence of binary spikes. As these spike activations can be very sparse, neuromorphic processors use this sparsity to improve energy efficiency. However, to close the accuracy gap to DNNs, more complex SNNs are required. The complexity of these new networks requires a scalable architecture with efficient neuromorphic cores. In this paper, we first compare and find the limitations of current neuromorphic architectures for large-scale convolutional SNNs. We also show an energy and throughput analysis of SPiking tiny Yolo-v4 (SPYv4), a state-of-the-art convolutional SNN, running on THOR, a state-of-the-art neuromorphic processor. Based on this analysis, we present a scalable digital architecture for convolutional SNNs. The proposed architecture is capable of running large convolutional SNNs like the SPYv4 network efficiently. On the proposed architecture, we compare eventbased scheduling with frame-based scheduling. By exploiting both activitation and weight sparsity, the architecture achieves an energy/SOP of 0.292 pJ/SOP, which is lower than state-of-the-art neuromorphic architectures.

Mixed-signal neuromorphic processors have brain-like organization and device physics optimized for emulation of spiking neural networks (SNNs), and offer an energy-efficient alternative for implementing artificial intelligence in applications where deep learning based on conventional digital computing is unfeasible or unsustainable. However, efficient use of such hardware requires appropriate configuration of its inhomogeneous, analog neurosynaptic circuits, with methods for sparse, spike-timing-based information encoding and processing. Furthermore, as neuromorphic processors are event-driven and asynchronous with massively parallel dynamic processing and colocated memory, they differ fundamentally from conventional von Neumann computers. Therefore, there is a need to investigate programming approaches and learning mechanisms for efficient use of neuromorphic processors, as well as abstractions required for large-scale integration of such devices into the present computational infrastructure of distributed digital systems. In this thesis, a disynaptic excitatory–inhibitory (E–I) element for resource-efficient generation of synaptic delay dynamics for spike-timing-based computation in neuromorphic hardware is proposed. Chip-in-the-loop experiments with a DYNAP-SE neuromorphic processor and SNN simulations are presented, demonstrating how such E–I elements leverage hardware inhomogeneity for representational variance and feature tuning for time-dependent pattern recognition. Using the E–I elements, spatiotemporal receptive fields with up to five dimensions per hardware neuron were characterized, for instance in a modified Spatiotemporal Correlator (STC) network and in an insect-inspired SNN. The energy dissipation of the proposed E–I element is one order of magnitude lower per lateral connection (0.65 vs. 9.6 nJ per spike) than the original delay-based hardware implementation of the STC. Thus, it is shown how the analog synaptic circuits could be used for efficient i

Deep learning based on neural networks emerged as a robust solution to various complex problems such as speech recognition and visual recognition. Deep learning relies on a great amount of iterative computation on a huge dataset. As we need to transfer a large amount of data and program between the CPU and the memory unit, the data transfer rate through a bus becomes a limiting factor for computing speed, which is known as Von Neumann bottleneck. Moreover, the data transfer between memory and computation spends a large amount of energy and cause significant delay. To overcome the limitation of Von Neumann bottleneck, neuromorphic computing with emerging nonvolatile memory (eNVM) devices has been proposed to perform iterative calculations in memory without transferring data to a processor. This dissertation presents energy efficient hardware implementation of neuromorphic computing applications using phase change memory (PCM), subquantum conductive bridge random access memory (CBRAM), Ag-based CBRAM, and CuOx-based resistive random access memory (RRAM). Although substantial progress has been made towards in-memory computing with synaptic devices, compact nanodevices implementing non-linear activation functions for efficient full-hardware implementation of deep neural networks is still missing. Since DNNs need to have a very large number of activations to achieve high accuracy, it is critical to develop energy and area efficient implementations of activation functions, which can be integrated on the periphery of the synaptic arrays. In this dissertation, we demonstrate a Mott activation neuron that implements the rectified linear unit function in the analogue domain. The integration of Mott activation neurons with a CBRAM crossbar array is also demonstrated in this dissertation.

Increasing complexity and data-generation rates in cyber-physical systems and the industrial Internet of things are calling for a corresponding increase in AI capabilities at the resource-constrained edges of the Internet. Meanwhile, the resource requirements of digital computing and deep learning are growing exponentially, in an unsustainable manner. One possible way to bridge this gap is the adoption of resource-efficient brain-inspired “neuromorphic” processing and sensing devices, which use event-driven, asynchronous, dynamic neurosynaptic elements with colocated memory for distributed processing and machine learning. However, since neuromorphic systems are fundamentally different from conventional von Neumann computers and clock-driven sensor systems, several challenges are posed to large-scale adoption and integration of neuromorphic devices into the existing distributed digital–computational infrastructure. Here, we describe the current landscape of neuromorphic computing, focusing on characteristics that pose integration challenges. Based on this analysis, we propose a microservice-based conceptual framework for neuromorphic systems integration, consisting of a neuromorphic-system proxy, which would provide virtualization and communication capabilities required in distributed systems of systems, in combination with a declarative programming approach offering engineering-process abstraction. We also present concepts that could serve as a basis for the realization of this framework, and identify directions for further research required to enable large-scale system integration of neuromorphic devices., Validerad;2023;Nivå 2;2023-02-17 (joosat);Funder: ECSEL JU (737 459); KDT JU (101007273); ERUF Interreg Nord, (NYPS 20202460)Licens fulltext: CC BY LicenseThis article has previously appeared as a manuscript in a thesis., Arrowhead Tools, DAIS, AI@Edge

This PhD dissertation explores the potential of two-dimensional (2D) layered materials to revolutionize computing and communication beyond today's electronics. The invention of metal-oxide-semiconductor field-effect-transistor (MOSFET) in the late 1959 has driven the need for dimensional scaling and new applications, making 2D materials with their exotic structural, physical, electrical, and magnetic properties, an attractive solution. This thesis will first present a comprehensive analysis of the various electrical contact topologies to 2D materials, leading to the development of the first comprehensive numerical model for contact resistance, addressing efficient charge injection. This work lays the foundation for the next stage of my research, which involves developing a comprehensive mobility model for quantifying transport in 2D materials to aid the design of high-performance 2D-FETs.By coupling the electrical charge injection and transport models, this thesis will discuss the results of the first comprehensive scaling analysis of 2D materials in emerging gate-all-around architectures, enabling the design of ultra-energy-efficient and compact transistors for sustaining Moore’s law. Further energy-efficiency gains can be achieved through use of steep-slope 2D-Tunneling Field-Effect-Transistors (2D-TFETs), and this thesis will then present the first compact model for such devices. This model is then used to benchmark 2D-TFET device performance in emerging neuromorphic (NM) computing architectures, designed with both neuronal and synaptic functionalities. The simulations promise impressive energy-efficiency gains of close to three orders of magnitude compared to state-of-the-art MOSFETs. In addition to the conventional charge-based computing methodologies, spin-based computing is another alternative computing platform that offers an additional route for energy-efficient computing. However, there are two major drawbacks to realizing useful spin-based computing devic

Neuromorphic sensors imitate the sparse and event-based communication seen in biological sensory organs and brains. Today's sensors can emit many millions of asynchronous events per second, which is challenging to process on conventional computers. To avoid bottleneck effects, there is a need to apply and improve concurrent and parallel processing of events. We present AEStream: A library to efficiently stream asynchronous events from inputs to outputs on conventional computers. AEStream leverages cooperative multitasking primitives known as coroutines to concurrently process individual events, which dramatically simplifies the integration with event-based peripherals, such as event-based cameras and (neuromorphic) asynchronous hardware. We explore the effects of coroutines in concurrent settings by benchmarking them against conventional threading mechanisms, and find that AEStream provides at least twice the throughput. We then apply AEStream in a real-Time edge detection task on a GPU and demonstrate 1.3 times faster processing with 5 times fewer memory operations., Part of ISBN 9781450399470QC 20230801

Machine learning algorithms are experiencing unprecedented attention, but their inherent computational complexity leads to high energy consumption. However, a paradigm shift in computing methods has the potential to address the issue. This shift could be a move towards analog in-memory computing, a method which uses Ohm’s and Kirchhoff’s Laws, and carries out the processing directly where data resides. This approach is being propelled by the development of memristors, versatile memory devices that are programmable and energy efficient. This thesis explores the capabilities of a newly engineered memristor device. This device, based on Ferroelectric Tunnel Junctions (FTJ), was developed by Lund University and presents promising technology for analog in-memory computing. In this thesis, the creation of a mathematical model took place within a simulated setting. This provided the foundation for a sensitivity analysis of chosen neural network algorithms operating on hardware featuring FTJ devices. A variety of techniques were deployed to mitigate the hardware imperfections, such as hardware-aware training, which enhanced the resilience of the algorithms. The outcomes from this investigative approach are promising, particularly regarding the inference processes in neural networks. Our research demonstrated the effectiveness of all applied mitigation techniques. The standout discovery was the robustness of the Transformer algorithm, compared to convolutional one, which proved capable of withstanding hardware imperfections while producing results on par with those of the digital model., Machine learning algorithms have become an integral part of our daily lives, powering applications and systems across various industries. However, their widespread use comes with a significant drawback - high power consumption. As the demand for machine learning continues to grow, there is an urgent need for innovative computing approaches that can tackle this power challenge. Enter Analog In-Memory Computing (AIMC), a promising solution that could revolutionize the way we process information. When combined with a type of memory cell called a Ferroelectric Tunnel Junction (FTJ) memristor, AIMC can perform the calculations needed for machine learning with much less power. In this study, we evaluated the effectiveness of FTJ memristors, engineered at NanoLund, in enabling AIMC. Our findings were encouraging. Interestingly, FTJ memristors demonstrated impressive performance when interfaced with the same kind of transformer neural network that powers ChatGPT, a state-of-the-art algorithm. But our work did not stop there. We also explored the potential of AIMC for training neural networks and designed several techniques to enhance the precision of our networks. Finding the proper way to map your application on the AIMC hardware is quite the challenge. Furthermore, we analyzed the sensitivity of the algortihms and proposed approaches that could make them functional. Through this comprehensive understanding, we suggest the implementation of AIMC for transformer models and pinpoint the key attributes for a successful design, known as mitigation techniques. The magic of AIMC lies in its simplicity. It's based on Ohm’s and Kirchhoff laws for multiplication and addition. These combined operations, known as multiplication accumulation (MAC), are extensively used for matrix operations required for machine learning algorithms. While the digital domain needs extensive hardware and power consumption to compute the operation, in the analog domain, simple resistors can do the trick.

Memristive crossbar arrays hold the great promise for fast and energy efficient neuromorphic computing due to their parallel data storage and processing capabilities. As the key component, memristor should achieve stable resistance switching (RS) characteristics with low energy inputs and be compatible with complementary metal–oxide–semiconductor (CMOS) technology. It should also exhibit sufficient device flexibility for applications in wearable electronics. In this thesis, we fabricate flexible memristors (FMs) based on Ag2S films, investigate their RS behavior and mechanism, demonstrate CMOS-compatible array integration and validate their computing applications. The thesis starts with a full-inorganic FM, utilizing ductile Ag2S thick films as both a flexible substrate and a functional electrolyte. The device exhibits dense multiple-level non-volatile states with a remarkable ON/OFF ratio of 106. The exceptional RS behavior is induced by sequential processes of Schottky barrier height (SBH) modification at the contact interface and silver filament formation inside the electrolyte. As a follow-up, we show that interface RS by SBH modification can be facilitated with smaller setting voltages. In contrast to traditional filamentary memristors, the sole interface RS achieves an ultralow switching energy of ~0.2 fJ. An image processing with interface RS indeed exhibits 2 orders of magnitude lower power than that with filamentary RS on the same hardware. Moreover, interface RS avoids the stochastic nature of filament formation and ablation inside electrolytes. The Ag2S-based FM operating with interface RS exhibits an impressive cycle-to-cycle variation of 1.4%, which is in direct contrast to the variation (28.9%) of filament RS extracted from the same device. Its significantly improved image learning ability over filament RS is also demonstrated during the frequent weight update process in simulations. Large-scale memristor array, with energy-efficient memristive units

The increasing demand for real-time, low-power, and intelligent systems has led to the development of edge computing, which involves processing data at the source directly from sensors rather than in the cloud or a centralized data center. Neuromorphic hardware, inspired by the structure and function of the biological nervous system, offers unique advantages for edge computing such as parallel processing, event-driven computation, and low power consumption with in memory computing. In memory computing enables the ability to perform online learning where learning can happen simultaneously with data input in real-time rather than the typical offline learning approach of presenting data in large batches for processing and then learning with gradient descent. With online learning, rapid adaptation to the unfamiliar becomes possible which is necessary for many applications such as user customization, and mobile robot navigation in unknown environments. Here we will demonstrate the machine learning algorithms we developed for use with neuromorphic hardware for online learning and rapid adaptation to unfamiliar data streamed from event-based sensors.The thesis begins by providing an overview of the state-of-the-art in edge computing and neuromorphic hardware, highlighting their strengths and limitations for machine learning. Next, we present our original contributions, including the development of new algorithms for learning on edge neuromorphic hardware, the implementation of these algorithms on state-of-the-art hardware platforms, and the evaluation of their performance in comparison to traditional machine learning algorithms.Our results demonstrate that edge neuromorphic hardware is a promising platform for machine learning, offering high performance, low latency, and low power consumption. Our algorithms are shown to be effective in real-world applications such as image classification, and mid-air gesture recognition. In conclusion, this thesis makes a significant cont

The brain, this 3-pound mass of tissue that can easily be held in one's palm, has an inherent computational complexity that has always inspired efforts to endorse machines with some of its remarkable characteristics. Ironically, the brain computes in its own way, compared to analog or digital computers, despite sharing key concepts with both. It employs analog computation but digital communication, through spikes, both of which improve robustness to noise. This unique combination defines a new computational paradigm that we have just started to explore. The reasons why neuromorphic systems are one of the fastest growing applications are not purely scientific, but mainly technological. For 50 years, the principle guiding computations has been Moore's Law, a macroscopic observation that we will always find ways to engineer faster, smaller and cheaper chips. But there are several reasons why Moore's Law is no longer able to keep up. First, physics: As we downsize transistors close to the atomic scale, it becomes difficult to regulate electron flow. Electrons do not necessarily adhere to Newtonian physics and may pass through the transistor barriers, a phenomenon called quantum tunneling. This makes our computer architectures inefficient. Second, we have long accepted the existence of a trade-off between computing faster and consuming less power, but this has never been a problem until we started approaching the physical limits of fabricating transistors. And the final nail in Moore's Law coffin is put by deep learning. Our computational needs are now orders of magnitude higher than what our systems can deliver. The von-Neumann paradigm is already having more than its fair share of inefficiency—to match its brilliance. And the reason is simple. Computers have been designed with feasibility, not efficiency, at their center. And nowhere are the effects of a bad design more imminent—or the opportunity for an alternative design more compelling – than in emerging technologie

We address the problem of testing artificial intelligence (AI) hardware accelerators implementing spiking neural networks (SNNs). We define a metric to quickly rank available samples for training and testing based on their fault detection capability. The metric measures the interclass spike count difference of a sample for the fault-free design. In particular, each sample is assigned a score equal to the spike count difference between the first two top classes. The hypothesis is that samples with small scores achieve high fault coverage because they are prone to misclassification, i.e., a small perturbation in the network due to a fault will result in these samples being misclassified with high probability. We show that the proposed metric correlates with the per-sample fault coverage and that retaining a set of high-ranked samples in the order of ten achieves near-perfect fault coverage for critical faults that affect the SNN accuracy. The proposed test generation approach is demonstrated on two SNNs modeled in Python and on actual neuromorphic hardware. We discuss fault modeling and perform an analysis to reduce the fault space so as to speed up test generation time. © 1982-2012 IEEE.

Magnons are elementary excitations in magnetic materials and undergo nonlinear multimode scattering processes at large input powers. In experiments and simulations, we show that the interaction between magnon modes of a confined magnetic vortex can be harnessed for pattern recognition. We study the magnetic response to signals comprising sine wave pulses with frequencies corresponding to radial mode excitations. Three-magnon scattering results in the excitation of different azimuthal modes, whose amplitudes depend strongly on the input sequences. We show that recognition rates above 95\% can be attained for four-symbol sequences using the scattered modes, with strong performance maintained with the presence of amplitude noise in the inputs.

In the ever-evolving landscape of robotics, the quest for advanced synthetic machines that seamlessly integrate with human lives and society becomes increasingly paramount. At the heart of this pursuit lies the intrinsic need for these machines to perceive, understand, and navigate their surroundings autonomously. Among the senses, vision emerges as a cornerstone of human perception, providing a wealth of information about the world we inhabit. Thus, it comes as no surprise that equipping robots with vision-based perception capabilities, or computer vision, has captivated researchers for decades. Recent breakthroughs, fueled by the advent of deep learning, have propelled computer vision to new heights. However, challenges persist in leveraging the power of deep learning, as its hunger for computational resources poses hurdles in the realm of robotics, particularly for small flying robots with their inherent limitations of payload and power consumption. This dissertation embarks on a journey that begins at the intersection of two groundbreaking technologies with the potential to revolutionize computer vision and enhance its accessibility to small robots: event-based cameras and neuromorphic processors. These two technologies draw inspiration from the information processing mechanisms employed by biological brains. Event-based cameras output sparse events encoding pixel-level brightness changes at microsecond resolution, while neuromorphic processors leverage the power of spiking neural networks to realize a sparse and asynchronous processing pipeline. Throughout this dissertation, comprehensive investigations have been conducted, presenting innovative solutions and advancements in the fields of computer vision and robotics. The thesis begins by presenting the winning solution of the 2019 AIRR autonomous drone racing competition, which showcases a monocular vision-based navigation approach specifically designed to address the limitations of conventio, Control & Simulation

This paper presents an end-to-end condition monitoring co-design model, from vibration measurement to anomaly detection, where conventional signal processing principles are combined with neuromorphic sensing and computing concepts to enable investigations of the potential improvements offered by brain-like information processing technologies. The use of machine learning in condition monitoring became increasingly popular for intelligent fault diagnosis in the last decade, taking advantage of the rapid developments in deep learning. However, the high computational cost of training and using deep neural networks prevents the use of such solutions for analysing the bulk of data generated by the resource constrained edge devices, i.e., the condition monitoring sensor systems, as only a minor fraction of data can be transmitted to the cloud or edge servers for analysis. There is an untapped potential to process this data and thereby improve intelligent fault diagnosis models using event-triggered sensing, spiking neural networks, and neuromorphic processors that substantially can improve the energy efficiency and capacity of embedded machine learning condition monitoring solutions. The proposed co-design model is evaluated on two use-cases involving rolling element bearing failures, one based on a labelled laboratory environment dataset, and one based on a wind turbine drivetrain bearing failure representing a real-world scenario with stochastic changes of machine state and unknown labels of the bearing condition. By adjusting co-design parameters, the resulting hybrid conventional/neuromorphic model show a comparable accuracy in detection performance for the laboratory dataset compared to the state-of-the-art reported in the literature. Similarly, for the wind turbine drivetrain dataset a bearing fault detection time comparable to that in previous work is obtained. This shows the successful implementation of a hybrid conventional/neuromorphic co-design model for conditi, Full text license: CC BY;ISBN for host publication: 978-1-936263-29-5

This data publication contains the data for our publication "Pattern recognition in reciprocal space with a magnon-scattering reservoir" published in Nature Communications. The dataset is structured in folders corresponding to the different figures in the paper. Folder Fig2 and Fig2 contain the experimental data measured with Brillouin-light-scattering microscopy. The files contain the data integrated for the measurement positions described in the methods section in a csv format. Forlder Fig4 contains the evaluated numerical data presented in the corresponding figure. The raw data generated with micromagnetic simulations is too large for this dataset and is available upon request by the authors.

Artificial neuronal devices are critical building blocks of neuromorphic computing systems and currently the subject of intense research motivated by application needs from new computing technology and more realistic brain emulation. Researchers have proposed a range of device concepts that can mimic neuronal dynamics and functions. Although the switching physics and device structures of these artificial neurons are largely different, their behaviors can be described by several neuron models in a more unified manner. In this paper, the reports of artificial neuronal devices based on emerging volatile switching materials are reviewed from the perspective of the demonstrated neuron models, with a focus on the neuronal functions implemented in these devices and the exploitation of these functions for computational and sensing applications. Furthermore, the neuroscience inspirations and engineering methods to enrich the neuronal dynamics that remain to be implemented in artificial neuronal devices and networks toward realizing the full functionalities of biological neurons are discussed.

Spiking neural network (SNN) is broadly deployed in neuromorphic devices to emulate brain function. In this context, SNN security becomes important while lacking in-depth investigation. To this end, we target the adversarial attack against SNNs and identify several challenges distinct from the artificial neural network (ANN) attack: 1) current adversarial attack is mainly based on gradient information that presents in a spatiotemporal pattern in SNNs, hard to obtain with conventional backpropagation algorithms; 2) the continuous gradient of the input is incompatible with the binary spiking input during gradient accumulation, hindering the generation of spike-based adversarial examples; and 3) the input gradient can be all-zeros (i.e., vanishing) sometimes due to the zero-dominant derivative of the firing function. Recently, backpropagation through time (BPTT)-inspired learning algorithms are widely introduced into SNNs to improve the performance, which brings the possibility to attack the models accurately given spatiotemporal gradient maps. We propose two approaches to address the above challenges of gradient-input incompatibility and gradient vanishing. Specifically, we design a gradient-to-spike (G2S) converter to convert continuous gradients to ternary ones compatible with spike inputs. Then, we design a restricted spike flipper (RSF) to construct ternary gradients that can randomly flip the spike inputs with a controllable turnover rate, when meeting all-zero gradients. Putting these methods together, we build an adversarial attack methodology for SNNs. Moreover, we analyze the influence of the training loss function and the firing threshold of the penultimate layer on the attack effectiveness. Extensive experiments are conducted to validate our solution. Besides the quantitative analysis of the influence factors, we also compare SNNs and ANNs against adversarial attacks under different attack methods. This work can help reveal what happens in SNN attacks and m

Photonic integrated circuits are facilitating the development of optical neural networks, which have the potential to be both faster and more energy efficient than their electronic counterparts since optical signals are especially well-suited for implementing matrix multiplications. However, accurate programming of photonic chips for optical matrix multiplication remains a difficult challenge. Here, we describe both simple analytical models and data-driven models for offline training of optical matrix multipliers. We train and evaluate the models using experimental data obtained from a fabricated chip featuring a Mach-Zehnder interferometer mesh implementing 3-by-3 matrix multiplication. The neural network-based models outperform the simple physics-based models in terms of prediction error. Furthermore, the neural network models are also able to predict the spectral variations in the matrix weights for up to 100 frequency channels covering the C-band. The use of neural network models for programming the chip for optical matrix multiplication yields increased performance on multiple machine learning tasks.

Neuromorphic electronics, being inspired by how the brain works, hold great promise to the successful implementation of smart artificial systems. Among several neuromorphic hardware issues, a robust device functionality under extreme temperature is of particular importance for practical applications. Given that the organic memristors for artificial synapse applications are demonstrated under room temperature, achieving a robust device performance at extremely low or high temperature is still utterly challenging. In this work, the temperature issue is addressed by tuning the functionality of the solution-based organic polymeric memristor. The optimized memristor demonstrates a reliable performance under both the cryogenic and high-temperature environments. The unencapsulated organic polymeric memristor shows a robust memristive response under test temperature ranging from 77 to 573 K. Utilizing X-ray photoelectron spectroscopy (XPS) and time-of-flight secondary-ion mass spectrometry (ToF-SIMS) depth profiling, the device working mechanism is unveiled by comparing the compositional profiles of the fresh and written organic polymeric memristors. A reversible ion migration induced by an applied voltage contributes to the characteristic switching behavior of the memristor. Herein, both the robust memristive response achieved at extreme temperatures and the verified device working mechanism will remarkably accelerate the development of memristors in neuromorphic systems. © 2023 Wiley-VCH GmbH.

Highly accurate classification methods for multi-task biomedical signal processing are reported, including neural networks. However, reported works are computationally expensive and power-hungry. Such bottlenecks make it hard to deploy existing approaches on edge platforms such as mobile and wearable devices. Gaining motivation from the good performance and high energy-efficiency of spiking neural networks (SNNs), a generic neuromorphic framework for edge healthcare and biomedical applications are proposed and evaluated on various tasks, including electroencephalography (EEG) based epileptic seizure prediction, electrocardiography (ECG) based arrhythmia detection, and electromyography (EMG) based hand gesture recognition. This approach, NeuroCARE, uses a unique sparse spike encoder to generate spike sequences from raw biomedical signals and makes classifications using the spike-based computing engine that combines the advantages of both CNN and SNN. An adaptive weight mapping method specifically co-designed with the spike encoder can efficiently convert CNN to SNN without performance deterioration. The evaluation results show that the overall performance, including the classification accuracy, sensitivity and F1 score, achieve 92.7, 96.7, and 85.7% for seizure prediction, arrhythmia detection and hand gesture recognition, respectively. In comparison with CNN topologies, the computation complexity is reduced by over 80.7% while the energy consumption and area occupation are reduced by over 80% and over 64.8%, respectively, indicating that the proposed neuromorphic computing approach is energy and area efficient and of high precision, which paves the way for deployment at edge platforms. Copyright © 2023 Tian, Yang, Zhao and Sawan.

Artificial intelligence (AI) is gaining strength, and materials science can both contribute to and profit from it. In a simultaneous progress race, new materials, systems, and processes can be devised and optimized thanks to machine learning (ML) techniques, and such progress can be turned into innovative computing platforms. Future materials scientists will profit from understanding how ML can boost the conception of advanced materials. This review covers aspects of computation from the fundamentals to directions taken and repercussions produced by computation to account for the origins, procedures, and applications of AI. ML and its methods are reviewed to provide basic knowledge of its implementation and its potential. The materials and systems used to implement AI with electric charges are finding serious competition from other information-carrying and processing agents. The impact these techniques have on the inception of new advanced materials is so deep that a new paradigm is developing where implicit knowledge is being mined to conceive materials and systems for functions instead of finding applications to found materials. How far this trend can be carried is hard to fathom, as exemplified by the power to discover unheard of materials or physical laws buried in data.

The rapid development in the field of artificial intelligence has increased the demand for neuromorphic computing hardware and its information-processing capability. A spintronics device is a promising candidate for neuromorphic computing hardware and can be used in extreme environments due to its high resistance to radiation. Improving the information-processing capability of neuromorphic computing is an important challenge for implementation. Herein, a novel neuromorphic computing framework using spintronics devices is proposed. This framework is called coupled spintronics reservoir (CSR) computing and exploits the high-dimensional dynamics of coupled spin-torque oscillators as a computational resource. The relationships among various bifurcations of the CSR and its information-processing capabilities through numerical experiments are analyzed and it is found that certain configurations of the CSR boost the information-processing capability of the spintronics reservoir toward or even beyond the standard level of machine learning networks. The effectiveness of our approach is demonstrated through conventional machine learning benchmarks and edge computing in real physical experiments using pneumatic artificial muscle-based wearables, which assist human operations in various environments. This study significantly advances the availability of neuromorphic computing for practical uses.

The rise of 5G and beyond systems has fuelled research in merging machine learning with wireless communications to achieve cognitive radios. However, the portability and limited power supply of radio frequency devices limits engineers' ability to combine them with powerful predictive models. This hinders the ability to support advanced 5G applications such as device-to-device (D2D) communication and dynamic spectrum sharing (DSS). This challenge has inspired a wave of research in energy efficient machine learning hardware with low computational and area overhead. In particular, hardware implementations of the delayed feedback reservoir (DFR) model show promising results for meeting these constraints while achieving high accuracy in cognitive radio applications. This thesis answers two research questions surrounding the applicability of FPGA DFR systems for DSS. First, can a DFR network implemented on an FPGA run faster and with lower power than a purely software approach? Second, can the system be implemented efficiently on an edge device running at less than 10 watts? Two systems are proposed that prove FPGA DFRs can achieve these feats: a mixed-signal circuit, followed by a high-level synthesis circuit. The implementations execute up to 58 times faster, and operate at more than 90% lower power than the software models. Furthermore, the lowest recorded average power of 0.130 watts proves that these approaches meet typical edge device constraints. When validated on the NARMA10 benchmark, the systems achieve a normalized error of 0.21 compared to state-of-the-art error values of 0.15. In a DSS task, the systems are able to predict spectrum occupancy with up to 0.87 AUC in high noise, multiple input, multiple output (MIMO) antenna configurations compared to 0.99 AUC in other works. At the end of this thesis, the trade-offs between the approaches are analyzed, and future directions for advancing this study are proposed.

A sustainable computing scenario demands more energy-efficient processors. Neuromorphic systems mimic biological functions by employing spiking neural networks for achieving brain-like efficiency, speed, adaptability, and intelligence. Current trends in neuromorphic technologies address the challenges of investigating novel materials, systems, and architectures for enabling high-integration and extreme low-power brain-inspired computing. This review collects the most recent trends in exploiting the physical properties of nonvolatile memory technologies for implementing efficient in-memory and in-device computing with spike-based neuromorphic architectures.

Organic mixed ionic–electronic conductors (OMIECs) are central to bioelectronic applications such as biosensors, health-monitoring devices, and neural interfaces, and have facilitated efficient next-generation brain-inspired computing and biohybrid systems. Despite these examples, smart and adaptive circuits that can locally process and optimize biosignals have not yet been realized. Here, a tunable sensing circuit is shown that can locally modulate biologically relevant signals like electromyograms (EMGs) and electrocardiograms (ECGs), that is based on a complementary logic inverter combined with a neuromorphic memory element, and that is constructed from a single polymer mixed conductor. It is demonstrated that a small neuromorphic array based on this material effects high classification accuracy in heartbeat anomaly detection. This high-performance material allows for straightforward monolithic integration, which reduces fabrication complexity while also achieving high on/off ratios with excellent ambient p- and n-type stability in transistor performance. This material opens a route toward simple and straightforward fabrication and integration of more sophisticated adaptive circuits for future smart bioelectronics.

Artificial intelligence (AI) has recently received significant attention as a key enabler for future 5G-and-beyond terrestrial wireless networks. The applications of AI to satellite communications is also gaining momentum to realize a more autonomous operation with reduced requirements in terms of human intervention. The adoption of AI for satellite communications will set new requirements on computing processors, which will need to support large workloads as efficiently as possible under harsh environmental conditions. In this context, neuromorphic processing (NP) is emerging as a bio-inspired solution to address pattern recognition tasks involving multiple, possibly unstructured, temporal signals and/or requiring continual learning. The key merits of the technology are energy efficiency and capacity for on-device adaptation. In this paper, we highlight potential use cases and applications of NP to satellite communications. We also explore major technical challenges for the implementation of space-based NP focusing on the available NP chipsets.

Hebbian learning based neural network learning rules when implemented on hardware, store their synaptic weights in the form of a two-dimensional matrix. The storage of synaptic weights demands large memory bandwidth and storage. While memory units are optimized for only row-wise memory access, Hebbian learning rules, like the spike-timing dependent plasticity, demand both row and column-wise access of memory. This dual pattern of memory access accounts for the dominant cost in terms of latency as well as energy for realization of large scale spiking neural networks in hardware. In order to reduce the memory access cost in Hebbian learning rules, a Column Update Elimination optimization has been previously implemented, with great efficacy, on the Bayesian Confidence Propagation neural network, that faces a similar challenge of dual pattern memory access. This thesis explores the possibility of extending the column update elimination optimization to spike-timing dependent plasticity, by simulating the learning rule on a two layer network of leaky integrate-and-fire neurons on an image classification task. The spike times are recorded for each neuron in the network, to derive a suitable probability distribution function for spike rates per neuron. This is then used to derive an ideal postsynaptic spike history buffer size for the given algorithm. The associated memory access reductions are analysed based on data to assess feasibility of the optimization to the learning rule., Hebbiansk inlärning baserat på neural nätverks inlärnings regler används vid implementering på hårdvara, de lagrar deras synaptiska vikter i form av en tvådimensionell matris. Lagringen av synaptiska vikter kräver stor bandbredds minne och lagring. Medan minnesenheter endast är optimerade för radvis minnesåtkomst. Hebbianska inlärnings regler kräver som spike-timing-beroende plasticitet, både rad- och kolumnvis åtkomst av minnet. Det dubbla mönstret av minnes åtkomsten står för den dominerande kostnaden i form av fördröjning såväl som energi för realiseringen av storskaliga spikande neurala nätverk i hårdvara. För att minska kostnaden för minnesåtkomst i hebbianska inlärnings regler har en Column Update Elimination-optimering tidigare implementerats, med god effektivitet på Bayesian Confidence Propagation neurala nätverket, som står inför en liknande utmaning med dubbel mönster minnesåtkomst. Denna avhandling undersöker möjligheten att utöka ColumnUpdate Elimination-optimeringen till spike-timing-beroende plasticitet. Detta genom att simulera inlärnings regeln på ett tvålagers nätverk av läckande integrera-och-avfyra neuroner på en bild klassificerings uppgift. Spike tiderna registreras för varje neuron i nätverket för att erhålla en lämplig sannolikhetsfördelning funktion för frekvensen av toppar per neuron. Detta används sedan för att erhålla en idealisk postsynaptisk spike historisk buffertstorlek för den angivna algoritmen. De associerade minnesåtkomst minskningarna analyseras baserat på data för att bedöma genomförbarheten av optimeringen av inlärnings regeln.

This dissertation is dedicated to using Memristive Spiking Neural Networks (MSNNs) for deep learning tasks such as image classification, visual associative memory tasks such as pattern recognition, and auditory cortex processing tasks such as sound localization (SL). The image classification model consists entirely of memristive neurons and memristive synapses utilizing deep learning plausible supervised learning rules. The pattern recognition fully MSNNs consists of memristive neurons and memristive synapses harnessing biologically plausible unsupervised learning rules. SL MSNNs emulate biological brain functionality with volatile memristive synapses.By developing a minimal circuit element memristive neuron -- Memristive Integrate-and-Fire (MIF) neuron -- with commercially accessible memristors, we are able to demonstrate large-scale fully MSNNs and apply the supervised Backpropagation Through Time (BPTT) algorithm to train networks, achieving state-of-the-art accuracy for several datasets. In addition, using a memristive unsupervised learning rule based on a continuously evolving alpha function membrane potential, we are also able to train large-scale fully MSNNs with Spiking-Time-Dependent Plasticity (STDP) and achieve high accuracies. Moreover, a volatile memristor is used for mimicking Short-Term Depression (STD) synapses such that we can simulate SL networks and pinpoint the direction of the sound coming from.The major contributions reported in this proposal include: • Development of the MIF neuron SPICE-level model; • Validation of the SPICE-level MIF neuron model to a Python-based simulation of large-scale MSNNs; • Simulation of a small-scale neuron network consisting of a presynaptic, a memristive synapse, and a postsynaptic MIF neuron to generate the STDP learning window; • Abstraction of memristive STDP with alpha functions; • Simulation of a large-scale, fully MSNN consisting of MIF neurons and memristive synapses for unsupervised learning in Python; • D

This paper introduces a new way of solving an edge user allocation problem. The problem is to be solved with a network of spiking neurons. This network should quickly and with low energy cost solve the optimization problem of allocating users to servers and minimizing the amount of servers hired to reduce the related hiring cost. The demonstrated method is a simulation of a method which could be implemented onto neuromorphic hardware. It is written in Python using the Brian2 spiking neural network simulator. The core of the method involves simulating an energy function through the use of circuit motifs. The dynamics of these circuit motifs mimic a search for the lowest energy point in an energy landscape, corresponding to a valid solution for the edge user allocation problem. The paper also shows the results of testing this network within the Brian2 environment.

Spiking neural networks using Leaky-Integrate-and-Fire (LIF) neurons and Spike-timing-depend Plasticity (STDP) learning, are commonly used as more biological possible networks. Compare to DNNs and RNNs, the LIF-STDP networks are models which are closer to the biological cortex. LIF-STDP neurons use spikes to communicate with each other, and they learn through the correlation among these pre- and post-synaptic spikes. Simulation of such networks usually requires high-performance supercomputers which are almost all based on von Neumann architecture that separates storage and computation. In von Neumann architecture solutions, memory access is the bottleneck even for highly optimized Application-Specific Integrated Circuits (ASICs). In this thesis, we propose an optimization method that can reduce the memory access cost by avoiding a dual-access pattern. In LIF-STDP networks, the weights usually are stored in the form of a two-dimensional matrix. Pre- and post-synaptic spikes trigger row and column access correspondingly. But this dual-access pattern is very costly for DRAM. We eliminate the column access by introducing a post-synaptic buffer and an approximation function. The post-synaptic spikes are recorded in the buffer and are processed at pre-synaptic spikes together with the row updates. This column update elimination method will introduce errors due to the limited buffer size. In our error analysis, the experiments show that the probability of introducing intolerable errors can be bounded to a very small number with proper buffer size and approximation function. We also present a performance analysis of the Column Update Elimination (CUE) optimization. The error analysis of the column updates elimination method is the main contribution of our work., Spikande neurala nätverk som använder LIF-neuroner och STDP-inlärning, används vanligtvis som ett mer biologiskt möjligt nätverk. Jämfört med DNN och RNN är LIF-STDP-nätverken modeller närmare den biologiska cortex. LIFSTDP-neuroner använder spikar för att kommunicera med varandra, och de lär sig genom korrelationen mellan dessa pre- och postsynaptiska spikar. Simulering av sådana nätverk kräver vanligtvis högpresterande superdatorer som nästan alla är baserade på von Neumann-arkitektur som separerar lagring och beräkning. I von Neumanns arkitekturlösningar är minnesåtkomst flaskhalsen även för högt optimerade Application-Specific Integrated Circuits (ASIC). I denna avhandling föreslår vi en optimeringsmetod som kan minska kostnaden för minnesåtkomst genom att undvika ett dubbelåtkomstmönster. I LIF-STDPnätverk lagras vikterna vanligtvis i form av en tvådimensionell matris. Preoch postsynaptiska toppar kommer att utlösa rad- och kolumnåtkomst på motsvarande sätt. Men detta mönster med dubbel åtkomst är mycket dyrt i DRAM. Vi eliminerar kolumnåtkomsten genom att införa en postsynaptisk buffert och en approximationsfunktion. De postsynaptiska topparna registreras i bufferten och bearbetas vid presynaptiska toppar tillsammans med raduppdateringarna. Denna metod för eliminering av kolumnuppdatering kommer att introducera fel på grund av den begränsade buffertstorleken. I vår felanalys visar experimenten att sannolikheten för att införa oacceptabla fel kan begränsas till ett mycket litet antal med korrekt buffertstorlek och approximationsfunktion. Vi presenterar också en prestandaanalys av CUE-optimeringen. Felanalysen av elimineringsmetoden för kolumnuppdateringar är det huvudsakliga bidraget från vårt arbete

Novel devices are being investigated as artificial synapse candidates for neuromorphic computing. These memory devices share the characteristics of an electronic element called memristor. The memristor can be regarded as a resistor with a history dependent resistance, which mimics the plasticity of a biological synapse. The present work presents various types of candidate devices that have been proposed in neuromorphic research, describes how they mimic a biological synapse and how they can be employed in artificial neuron network architectures.

Synaptic devices with linear high-speed switching can accelerate learning in artificial neural networks (ANNs) embodied in hardware. Conventional resistive memories however suffer from high write noise and asymmetric conductance tuning, preventing parallel programming of ANN arrays. Electrochemical random-access memories (ECRAMs), where resistive switching occurs by ion insertion into a redox-active channel, aim to address these challenges due to their linear switching and low noise. ECRAMs using 2D materials and metal oxides however suffer from slow ion kinetics, whereas organic ECRAMs enable high-speed operation but face challenges toward on-chip integration due to poor temperature stability of polymers. Here, ECRAMs using 2D titanium carbide (Ti3C2Tx) MXene that combine the high speed of organics and the integration compatibility of inorganic materials in a single high-performance device are demonstrated. These ECRAMs combine the speed, linearity, write noise, switching energy, and endurance metrics essential for parallel acceleration of ANNs, and importantly, they are stable after heat treatment needed for back-end-of-line integration with Si electronics. The high speed and performance of these ECRAMs introduces MXenes, a large family of 2D carbides and nitrides with more than 30 stoichiometric compositions synthesized to date, as promising candidates for devices operating at the nexus of electrochemistry and electronics., QC 20220615