A ideia de usar redes neurais biologicamente inspiradas na computação tem sido amplamente utilizada nas últimas décadas. O fato essencial neste paradigma é que um neurônio pode integrar e processar informações, e esta informação pode ser revelada por sua atividade de pulsos. Ao descrever a dinâmica de um único neurônio usando um modelo matemático, uma rede pode ser implementada utilizando um conjunto desses neurônios, onde a atividade pulsante de cada neurônio irá conter contribuições, ou informações, da atividade pulsante da rede em que está inserido. Neste trabalho é apresentado um controlador de posição no eixo Z utilizando fusão de sensores baseado no paradigma de Redes Neurais Recorrentes. O sistema proposto utiliza uma Máquina de Estado Líquido (LSM) para controlar o robô colaborativo BAXTER. O framework foi projetado para trabalhar em paralelo com as LSMs que executam trajetórias em formas fechadas de duas dimensões, com o objetivo de manter uma caneta de feltro em contato com a superfície de desenho, dados de sensores de força e distância são alimentados ao controlador. O sistema foi treinado utilizando dados de um controlador Proporcional Integral Derivativo (PID), fundindo dados de ambos sensores. Resultados mostram que a LSM foi capaz de aprender o comportamento do controlador PID em diferentes situações. / The idea of employing biologically inspired neural networks to perform computation has been widely used over the last decades. The essential fact in this paradigm is that a neuron can integrate and process information, and this information can be revealed by its spiking activity. By describing the dynamics of a single neuron using a mathematical model, a network in which the spiking activity of every single neuron will get contributions, or information, from the spiking activity of the embedded network. A positioning controller based on Spiking Neural Networks for sensor fusion suitable to run on a neuromorphic computer is presented in this work. The proposed framework uses the paradigm of reservoir computing to control the collaborative robot BAXTER. The system was designed to work in parallel with Liquid State Machines that performs trajectories in 2D closed shapes. In order to keep a felt pen touching a drawing surface, data from sensors of force and distance are fed to the controller. The system was trained using data from a Proportional Integral Derivative controller, merging the data from both sensors. The results show that the LSM can learn the behavior of a PID controller on di erent situations.

SpiNNaker is an unconventional supercomputer architecture designed to simulate up to one billion biologically realistic neurons in real-time. To achieve this goal, SpiNNaker employs a novel network architecture which poses a number of practical problems in scaling up from desktop prototypes to machine room filling installations. SpiNNaker's hexagonal torus network topology has received mostly theoretical treatment in the literature. This thesis tackles some of the challenges encountered when building `real-world' systems. Firstly, a scheme is devised for physically laying out hexagonal torus topologies in machine rooms which avoids long cables; this is demonstrated on a half-million core SpiNNaker prototype. Secondly, to improve the performance of existing routing algorithms, a more efficient process is proposed for finding (logically) short paths through hexagonal torus topologies. This is complemented by a formula which provides routing algorithms with greater flexibility when finding paths, potentially resulting in a more balanced network utilisation. The scale of SpiNNaker's network and the models intended for it also present their own challenges. Placement and routing algorithms are developed which assign processes to nodes and generate paths through SpiNNaker's network. These algorithms minimise congestion and tolerate network faults. The proposed placement algorithm is inspired by techniques used in chip design and is shown to enable larger applications to run on SpiNNaker than the previous state-of-the-art. Likewise the routing algorithm developed is able to tolerate network faults, inevitably present in large-scale systems, with little performance overhead.

Neuromorphic information processing systems mimic the dynamics of neurons and synapses, and the architecture of biological nervous systems. By using a combination of sub-threshold analog circuits, and fast programmable digital circuits, spiking neural networks with co-localized memory and computation can be implemented, enabling more energy-efficient information processing than conventional von Neumann digital computers. When configuring such a spiking neural network, the variability caused by device mismatch of the analog electronic circuits must be managed and exploited. While pre-trained spiking neural networks have been approximated in neuromorphic processors in previous work, configuration methods and tools need to be developed that make efficient use of the high number of inhomogeneous analog neuron and synapse circuits in a systematic manner. The aim of the work presented here is to investigate such automatic configuration methods, focusing in particular on Monte Carlo methods, and to develop software for training and configuration of the Dynap-SE neuromorphic processor, which is based on the Dynamic Neuromorphic Asynchronous Processor (DYNAP) architecture. A Monte Carlo optimization method enabling configuration of spiking neural networks on the Dynap-SE is developed and tested with the Metropolis-Hastings algorithm in the low-temperature limit. The method is based on a hardware-in-the-loop setup where a PC performs online optimization of a Dynap-SE, and the resulting system is tested by reproducing properties of small neural networks in the auditory system of field crickets. It is shown that the system successfully configures two different auditory neural networks, consisting of three and four neurons respectively. However, appropriate bias parameter values defining the dynamic properties of the analog neuron and synapse circuits must be manually defined prior to optimization, which is time consuming and should be included in the optimization protocol in future work.

A ideia de usar redes neurais biologicamente inspiradas na computação tem sido amplamente utilizada nas últimas décadas. O fato essencial neste paradigma é que um neurônio pode integrar e processar informações, e esta informação pode ser revelada por sua atividade de pulsos. Ao descrever a dinâmica de um único neurônio usando um modelo matemático, uma rede pode ser implementada utilizando um conjunto desses neurônios, onde a atividade pulsante de cada neurônio irá conter contribuições, ou informações, da atividade pulsante da rede em que está inserido. Neste trabalho é apresentado um controlador de posição no eixo Z utilizando fusão de sensores baseado no paradigma de Redes Neurais Recorrentes. O sistema proposto utiliza uma Máquina de Estado Líquido (LSM) para controlar o robô colaborativo BAXTER. O framework foi projetado para trabalhar em paralelo com as LSMs que executam trajetórias em formas fechadas de duas dimensões, com o objetivo de manter uma caneta de feltro em contato com a superfície de desenho, dados de sensores de força e distância são alimentados ao controlador. O sistema foi treinado utilizando dados de um controlador Proporcional Integral Derivativo (PID), fundindo dados de ambos sensores. Resultados mostram que a LSM foi capaz de aprender o comportamento do controlador PID em diferentes situações. / The idea of employing biologically inspired neural networks to perform computation has been widely used over the last decades. The essential fact in this paradigm is that a neuron can integrate and process information, and this information can be revealed by its spiking activity. By describing the dynamics of a single neuron using a mathematical model, a network in which the spiking activity of every single neuron will get contributions, or information, from the spiking activity of the embedded network. A positioning controller based on Spiking Neural Networks for sensor fusion suitable to run on a neuromorphic computer is presented in this work. The proposed framework uses the paradigm of reservoir computing to control the collaborative robot BAXTER. The system was designed to work in parallel with Liquid State Machines that performs trajectories in 2D closed shapes. In order to keep a felt pen touching a drawing surface, data from sensors of force and distance are fed to the controller. The system was trained using data from a Proportional Integral Derivative controller, merging the data from both sensors. The results show that the LSM can learn the behavior of a PID controller on di erent situations.

Ces travaux de thèse, réalisés dans une équipe concevant des circuits analogiques neuromimétiques suivant le modèle d’Hodgkin-Huxley, concernent la modélisation de neurones biologiques, plus précisément, l’estimation des paramètres de modèles de neurones. Une première partie de ce manuscrit s’attache à faire le lien entre la modélisation neuronale et l’optimisation. L’accent est mis sur le modèle d’Hodgkin- Huxley pour lequel il existait déjà une méthode d’extraction des paramètres associée à une technique de mesures électrophysiologiques (le voltage-clamp) mais dont les approximations successives rendaient impossible la détermination précise de certains paramètres. Nous proposons dans une seconde partie une méthode alternative d’estimation des paramètres du modèle d’Hodgkin-Huxley s’appuyant sur l’algorithme d’évolution différentielle et qui pallie les limitations de la méthode classique. Cette alternative permet d’estimer conjointement tous les paramètres d’un même canal ionique. Le troisième chapitre est divisé en trois sections. Dans les deux premières, nous appliquons notre nouvelle technique à l’estimation des paramètres du même modèle à partir de données biologiques, puis développons un protocole automatisé de réglage de circuits neuromimétiques, canal ionique par canal ionique. La troisième section présente une méthode d’estimation des paramètres à partir d’enregistrements de la tension de membrane d’un neurone, données dont l’acquisition est plus aisée que celle des courants ioniques. Le quatrième et dernier chapitre, quant à lui, est une ouverture vers l’utilisation de petits réseaux d’une centaine de neurones électroniques : nous réalisons une étude logicielle de l’influence des propriétés intrinsèques de la cellule sur le comportement global du réseau dans le cadre des oscillations gamma. / These works, which were conducted in a research group designing neuromimetic integrated circuits based on the Hodgkin-Huxley model, deal with the parameter estimation of biological neuron models. The first part of the manuscript tries to bridge the gap between neuron modeling and optimization. We focus our interest on the Hodgkin-Huxley model because it is used in the group. There already existed an estimation method associated to the voltage-clamp technique. Nevertheless, this classical estimation method does not allow to extract precisely all parameters of the model, so in the second part, we propose an alternative method to jointly estimate all parameters of one ionic channel avoiding the usual approximations. This method is based on the differential evolution algorithm. The third chaper is divided into three sections : the first two sections present the application of our new estimation method to two different problems, model fitting from biological data and development of an automated tuning of neuromimetic chips. In the third section, we propose an estimation technique using only membrane voltage recordings – easier to mesure than ionic currents. Finally, the fourth and last chapter is a theoretical study preparing the implementation of small neural networks on neuromimetic chips. More specifically, we try to study the influence of cellular intrinsic properties on the global behavior of a neural network in the context of gamma oscillations.

Ces travaux ont été menés dans le cadre du projet européen FACETS-ITN. Nous avons contribué à la simulation de cellules corticales grâce à des données expérimentales d'électrophysiologie comme référence et d'un circuit intégré neuromorphique comme simulateur. Les propriétés intrinsèques temps réel de nos circuits neuromorphiques à base de modèles à conductance, autorisent une exploration détaillée des différents types de neurones. L'aspect analogique des circuits intégrés permet le développement d'un simulateur matériel temps réel à l'échelle du réseau. Le deuxième objectif de cette thèse est donc de contribuer au développement d'une plate-forme mixte - matérielle et logicielle - dédiée à la simulation de réseaux de neurones impulsionnels. / This work has been supported by the European FACETS-ITN project. Within the frameworkof this project, we contribute to the simulation of cortical cell types (employingexperimental electrophysiological data of these cells as references), using a specific VLSIneural circuit to simulate, at the single cell level, the models studied as references in theFACETS project. The real-time intrinsic properties of the neuromorphic circuits, whichprecisely compute neuron conductance-based models, will allow a systematic and detailedexploration of the models, while the physical and analog aspect of the simulations, as opposedthe software simulation aspect, will provide inputs for the development of the neuralhardware at the network level. The second goal of this thesis is to contribute to the designof a mixed hardware-software platform (PAX), specifically designed to simulate spikingneural networks. The tasks performed during this thesis project included: 1) the methodsused to obtain the appropriate parameter sets of the cortical neuron models that can beimplemented in our analog neuromimetic chip (the parameter extraction steps was validatedusing a bifurcation analysis that shows that the simplified HH model implementedin our silicon neuron shares the dynamics of the HH model); 2) the fully customizablefitting method, in voltage-clamp mode, to tune our neuromimetic integrated circuits usinga metaheuristic algorithm; 3) the contribution to the development of the PAX systemin terms of software tools and a VHDL driver interface for neuron configuration in theplatform. Finally, it also addresses the issue of synaptic tuning for future SNN simulation.

I extended an artificial life platform called GReaNs (the name stands for Gene Regulatory evolving artificial Networks) to explore the evolutionary abilities of biologically inspired Spiking Neural Network (SNN) model. The encoding of SNNs in GReaNs was inspired by the encoding of gene regulatory networks. As proof-of-principle, I used GReaNs to evolve SNNs to obtain a network with an output neuron which generates a predefined spike train in response to a specific input. Temporal pattern recognition was one of the main tasks during my studies. It is widely believed that nervous systems of biological organisms use temporal patterns of inputs to encode information. The learning technique used for temporal pattern recognition is not clear yet. I studied the ability to evolve spiking networks with different numbers of interneurons in the absence and the presence of noise to recognize predefined temporal patterns of inputs. Results showed, that in the presence of noise, it was possible to evolve successful networks. However, the networks with only one interneuron were not robust to noise. The foraging behaviour of many small animals depends mainly on their olfactory system. I explored whether it was possible to evolve SNNs able to control an agent to find food particles on 2-dimensional maps. Using ring rate encoding to encode the sensory information in the olfactory input neurons, I managed to obtain SNNs able to control an agent that could detect the position of the food particles and move toward it. Furthermore, I did unsuccessful attempts to use GReaNs to evolve an SNN able to control an agent able to collect sound sources from one type out of several sound types. Each sound type is represented as a pattern of different frequencies. In order to use the computational power of neuromorphic hardware, I integrated GReaNs with the SpiNNaker hardware system. Only the simulation part was carried out using SpiNNaker, but the rest steps of the genetic algorithm were done with GReaNs.

Artificial neural network simulators are a research field which attracts the interest of researchers from various fields, from biology to computer science. The final objectives are the understanding of the mechanisms underlying the human brain, how to reproduce them in an artificial environment, and how drugs interact with them. Multiple neural models have been proposed, each with their peculiarities, from the very complex and biologically realistic Hodgkin-Huxley neuron model to the very simple 'leaky integrate-and-fire' neuron. However, despite numerous attempts to understand the learning behaviour of the synapses, few models have been proposed. Spike-Timing-Dependent Plasticity (STDP) is one of the most relevant and biologically plausible models, and some variants (such as the triplet-based STDP rule) have been proposed to accommodate all biological observations. The research presented in this thesis focuses on a novel learning rule, based on the spike-pair STDP algorithm, which provides a statistical approach with the advantage of being less computationally expensive than the standard STDP rule, and is therefore suitable for its implementation on stand-alone computational units. The environment in which this research work has been carried out is the SpiNNaker project, which aims to provide a massively parallel computational substrate for neural simulation. To support such research, two other topics have been addressed: the first is a way to inject spikes into the SpiNNaker system through a non-real-time channel such as the Ethernet link, synchronising with the timing of the SpiNNaker system. The second research topic is focused on a way to route spikes in the SpiNNaker system based on populations of neurons. The three topics are presented in sequence after a brief introduction to the SpiNNaker project. Future work could include structural plasticity (also known as synaptic rewiring); here, during the simulation of neural networks on the SpiNNaker system, axons, dendrites and synapses may be grown or pruned according to biological observations.

<p>Deep Neural Networks (DNNs) have greatly advanced the state-of-the-art in a wide range of machine learning tasks involving image, video, speech and text analytics, and are deployed in numerous widely-used products and services. Improvements in the capabilities of hardware platforms such as Graphics Processing Units (GPUs) and specialized accelerators have been instrumental in enabling these advances as they have allowed more complex and accurate networks to be trained and deployed. However, the enormous computational and memory demands of DNNs continue to increase with growing data size and network complexity, posing a continuing challenge to computing system designers. For instance, state-of-the-art image recognition DNNs require hundreds of millions of parameters and hundreds of billions of multiply-accumulate operations while state-of-the-art language models require hundreds of billions of parameters and several trillion operations to process a single input instance. Another major obstacle in the adoption of DNNs, despite their impressive accuracies on a range of datasets, has been their lack of robustness. Specifically, recent efforts have demonstrated that small, carefully-introduced input perturbations can force a DNN to behave in unexpected and erroneous ways, which can have to severe consequences in several safety-critical DNN applications like healthcare and autonomous vehicles. In this dissertation, we explore approximate computing as an avenue to improve the speed and energy efficiency of DNNs, as well as their robustness to input perturbations.</p> <p> </p> <p>Approximate computing involves executing selected computations of an application in an approximate manner, while generating favorable trade-offs between computational efficiency and output quality. The intrinsic error resilience of machine learning applications makes them excellent candidates for approximate computing, allowing us to achieve execution time and energy reductions with minimal effect on the quality of outputs. This dissertation performs a comprehensive analysis of different approximate computing techniques for improving the execution efficiency of DNNs. Complementary to generic approximation techniques like quantization, it identifies approximation opportunities based on the specific characteristics of three popular classes of networks - Feed-forward Neural Networks (FFNNs), Recurrent Neural Networks (RNNs) and Spiking Neural Networks (SNNs), which vary considerably in their network structure and computational patterns.</p> <p> </p> <p>First, in the context of feed-forward neural networks, we identify sparsity, or the presence of zero values in the data structures (activations, weights, gradients and errors), to be a major source of redundancy and therefore, an easy target for approximations. We develop lightweight micro-architectural and instruction set extensions to a general-purpose processor core that enable it to dynamically detect zero values when they are loaded and skip future instructions that are rendered redundant by them. Next, we explore LSTMs (the most widely used class of RNNs), which map sequences from an input space to an output space. We propose hardware-agnostic approximations that dynamically skip redundant symbols in the input sequence and discard redundant elements in the state vector to achieve execution time benefits. Following that, we consider SNNs, which are an emerging class of neural networks that represent and process information in the form of sequences of binary spikes. Observing that spike-triggered updates along synaptic connections are the dominant operation in SNNs, we propose hardware and software techniques to identify connections that can be minimally impact the output quality and deactivate them dynamically, skipping any associated updates.</p> <p> </p> <p>The dissertation also delves into the efficacy of combining multiple approximate computing techniques to improve the execution efficiency of DNNs. In particular, we focus on the combination of quantization, which reduces the precision of DNN data-structures, and pruning, which introduces sparsity in them. We observe that the ability of pruning to reduce the memory demands of quantized DNNs decreases with precision as the overhead of storing non-zero locations alongside the values starts to dominate in different sparse encoding schemes. We analyze this overhead and the overall compression of three different sparse formats across a range of sparsity and precision values and propose a hybrid compression scheme that identifies that optimal sparse format for a pruned low-precision DNN.</p> <p> </p> <p>Along with improved execution efficiency of DNNs, the dissertation explores an additional advantage of approximate computing in the form of improved robustness. We propose ensembles of quantized DNN models with different numerical precisions as a new approach to increase robustness against adversarial attacks. It is based on the observation that quantized neural networks often demonstrate much higher robustness to adversarial attacks than full precision networks, but at the cost of a substantial loss in accuracy on the original (unperturbed) inputs. We overcome this limitation to achieve the best of both worlds, i.e., the higher unperturbed accuracies of the full precision models combined with the higher robustness of the low precision models, by composing them in an ensemble.</p> <p> </p> <p><br></p><p>In summary, this dissertation establishes approximate computing as a promising direction to improve the performance, energy efficiency and robustness of neural networks.</p>

<div>Deep learning models have become the de-facto solution in various fields like computer vision, natural language processing, robotics, drug discovery, and many others. The skyrocketing performance and success of multi-layer neural networks comes at a significant power and energy cost. Thus, there is a need to rethink the current trajectory and explore different computing frameworks. One such option is spiking neural networks (SNNs) that is inspired from the spike-based processing observed in biological brains. SNNs operating with binary signals (or spikes), can potentially be an energy-efficient alternative to the power-hungry analog neural networks (ANNs) that operate on real-valued analog signals. The binary all-or-nothing spike-based communication in SNNs implemented on event-driven hardware offers a low-power alternative to ANNs. A spike is a Delta function with magnitude 1. With all its appeal for low power, training SNNs efficiently for high accuracy remains an active area of research. The existing ANN training methodologies when applied to SNNs, results in networks that have very high latency. Supervised training of SNNs with spikes is challenging (due to discontinuous gradients) and resource-intensive (time, compute, and memory).Thus, we propose compression methods, training methodologies, learning rules</div><div><br></div><div>First, we propose compression techniques for SNNs based on unsupervised spike timing dependent plasticity (STDP) model. We present a sparse SNN topology where non-critical connections are pruned to reduce the network size and the remaining critical synapses are weight quantized to accommodate for limited conductance levels in emerging in-memory computing hardware . Pruning is based on the power law weight-dependent</div><div>STDP model; synapses between pre- and post-neuron with high spike correlation are retained, whereas synapses with low correlation or uncorrelated spiking activity are pruned. The process of pruning non-critical connections and quantizing the weights of critical synapses is</div><div>performed at regular intervals during training.</div><div><br></div><div>Second, we propose a multimodal SNN that combines two modalities (image and audio). The two unimodal ensembles are connected with cross-modal connections and the entire network is trained with unsupervised learning. The network receives inputs in both modalities for the same class and</div><div>predicts the class label. The excitatory connections in the unimodal ensemble and the cross-modal connections are trained with STDP. The cross-modal connections capture the correlation between neurons of different modalities. The multimodal network learns features of both modalities and improves the classification accuracy compared to unimodal topology, even when one of the modality is distorted by noise. The cross-modal connections are only excitatory and do not inhibit the normal activity of the unimodal ensembles. </div><div><br></div><div>Third, we explore supervised learning methods for SNNs.Many works have shown that an SNN for inference can be formed by copying the weights from a trained ANN and setting the firing threshold for each layer as the maximum input received in that layer. These type of converted SNNs require a large number of time steps to achieve competitive accuracy which diminishes the energy savings. The number of time steps can be reduced by training SNNs with spike-based backpropagation from scratch, but that is computationally expensive and slow. To address these challenges, we present a computationally-efficient training technique for deep SNNs. We propose a hybrid training methodology:</div><div>1) take a converted SNN and use its weights and thresholds as an initialization step for spike-based backpropagation, and 2) perform incremental spike-timing dependent backpropagation (STDB) on this carefully initialized network to obtain an SNN that converges within few epochs and requires fewer time steps for input processing. STDB is performed with a novel surrogate gradient function defined using neuron’s spike time. The weight update is proportional to the difference in spike timing between the current time step and the most recent time step the neuron generated an output spike.</div><div><br></div><div>Fourth, we present techniques to further reduce the inference latency in SNNs. SNNs suffer from high inference latency, resulting from inefficient input encoding, and sub-optimal settings of the neuron parameters (firing threshold, and membrane leak). We propose DIET-SNN, a low-latency deep spiking network that is trained with gradient descent to optimize the membrane leak and the firing threshold along with other network parameters (weights). The membrane leak and threshold for each layer of the SNN are optimized with end-to-end backpropagation to achieve competitive accuracy at reduced latency. The analog pixel values of an image are directly applied to the input layer of DIET-SNN without the need to convert to spike-train. The first convolutional layer is trained to convert inputs into spikes where leaky-integrate-and-fire (LIF) neurons integrate the weighted inputs and generate an output spike when the membrane potential crosses the trained firing threshold. The trained membrane leak controls the flow of input information and attenuates irrelevant inputs to increase the activation sparsity in the convolutional and dense layers of the network. The reduced latency combined with high activation sparsity provides large improvements in computational efficiency.</div><div><br></div><div>Finally, we explore the application of SNNs in sequential learning tasks. We propose LITE-SNN, a lightweight SNN suitable for sequential learning tasks on data from dynamic vision sensors (DVS) and natural language processing (NLP). In general sequential data is processed with complex recurrent neural networks (like long short-term memory (LSTM), and gated recurrent unit (GRU)) with explicit feedback connections and internal states to handle the long-term dependencies. Whereas neuron models in SNNs - integrate-and-fire (IF) or leaky-integrate-and-fire (LIF) - have implicit feedback in their internal state (membrane potential) by design and can be leveraged for sequential tasks. The membrane potential in the IF/LIF neuron integrates the incoming current and outputs an event (or spike) when the potential crosses a threshold value. Since SNNs compute with highly sparse spike-based spatio-temporal data, the energy/inference is lower than LSTMs/GRUs. SNNs also have fewer parameters than LSTM/GRU resulting in smaller models and faster inference. We observe the problem of vanishing gradients in vanilla SNNs for longer sequences and implement a convolutional SNN with attention layers to perform sequence-to-sequence learning tasks. The inherent recurrence in SNNs, in addition to the fully parallelized convolutional operations, provides an additional mechanism to model sequential dependencies and leads to better accuracy than convolutional neural networks with ReLU activations.</div>