Due to the limited computation and storage resources of Internet of Things (IoT) devices, many emerging intelligent applications based on deep learning techniques heavily depend on cloud computing for computation and storage. However, cloud computing faces technical issues with long latency, poor reliability, and weak privacy, resulting in the need for on-device computation and storage. Also, on-device computation is essential for many time-critical applications, which require real-time data processing and energy-efficient. Furthermore, the escalating requirements for on-device processing are driven by network bandwidth limitations and consumer anticipations concerning data privacy and user experience. In the realm of computing, there is a growing interest in exploring novel technologies that can facilitate ongoing advancements in performance. Of the various prospective avenues, the field of neuromorphic computing has garnered significant recognition as a crucial means to achieve fast and energy-efficient machine intelligence applications for IoT devices. The programming of neuromorphic computing hardware typically involves the construction of a spiking neural network (SNN) capable of being deployed onto the designated neuromorphic hardware. This dissertation presents a range of methodologies aimed at enhancing the precision and energy efficiency of SNNs. To be more precise, these advancements are achieved by incorporating four essential methods. The first method is the quantization of neural networks through knowledge distillation. This work introduces a quantization technique that effectively reduces the computational and storage resource requirements of a model while minimizing the loss of accuracy. To further enhance the reduction of quantization errors, the second method introduces a novel quantization-aware training algorithm specifically designed for training quantized spiking neural network (SNN) models intended for execution on the Loihi chip, a specialized neuromorphic computing chip. SNNs generally exhibit lower accuracy performance compared to deep neural networks (DNNs). The third approach introduces a DNN-SNN co-learning algorithm, which enhances the performance of SNN models by leveraging knowledge obtained from DNN models. The design of the neural architecture plays a vital role in enhancing the accuracy and energy efficiency of an SNN model. The fourth method presents a novel neural architecture search algorithm specifically tailored for SNNs on the Loihi chip. The method selects an optimal architecture based on gradients induced by the architecture at initialization across different data samples without the need for training the architecture. To demonstrate the effectiveness and performance across diverse machine intelligence applications, our methods are evaluated through (i) image classification, (ii) spectrum sensing, and (iii) modulation symbol detection. / Doctor of Philosophy / In the emerging Internet of Things (IoT), our everyday devices, from smart home gadgets to wearables, can autonomously make intelligent decisions. However, due to their limited computing power and storage, many IoT devices heavily depend on cloud computing, which brings along issues like slow response times, privacy concerns, and unreliable connections. Neuromorphic computing is a recognized and crucial approach for achieving fast and energy-efficient machine intelligence applications in IoT devices. Inspired by the human brain's neural networks, this cutting-edge approach allows devices to perform complex tasks efficiently and in real-time. The programming of this neuromorphic hardware involves creating spiking neural networks (SNNs). This dissertation presents several innovative methods to improve the precision and energy efficiency of these SNNs. Firstly, a technique called "quantization" reduces the computational and storage requirements of models without sacrificing accuracy. Secondly, a unique training algorithm is designed to enhance the performance of SNN models. Thirdly, a clever co-learning algorithm allows SNN models to learn from traditional deep neural networks (DNNs), further improving their accuracy. Lastly, a novel neural architecture search algorithm finds the best architecture for SNNs on the designated neuromorphic chip, without the need for extensive training. By making IoT devices smarter and more efficient, neuromorphic computing brings us closer to a world where our gadgets can perform intelligent tasks independently, enhancing convenience and privacy for users across the globe.

Humans possess the fascinating ability to communicate the most complex of ideas through spoken language, without requiring any external tools. This process has two sides—a speaker producing speech, and a listener comprehending it. While the two actions are intertwined in many ways, they entail differential activation of neural circuits in the brains of the speaker and the listener. Both processes are the active subject of artificial intelligence research, under the names of speech synthesis and automatic speech recognition, respectively. While the capabilities of these artificial models are approaching human levels, there are still many unanswered questions about how our brains do this task effortlessly. But the advances in these artificial models allow us the opportunity to study human speech recognition through a computational lens that we did not have before. This dissertation explores the intricate processes of speech perception and comprehension by drawing parallels between artificial and biological neural networks, through the use of computational frameworks that attempt to model either the brain circuits involved in speech recognition, or the process of speech recognition itself. There are two general types of analyses in this dissertation. The first type involves studying neural responses recorded directly through invasive electrophysiology from human participants listening to speech excerpts. The second type involves analyzing artificial neural networks trained to perform the same task of speech recognition, as a potential model for our brains. The first study introduces a novel framework leveraging deep neural networks (DNNs) for interpretable modeling of nonlinear sensory receptive fields, offering an enhanced understanding of auditory neural responses in humans. This approach not only predicts auditory neural responses with increased accuracy but also deciphers distinct nonlinear encoding properties, revealing new insights into the computational principles underlying sensory processing in the auditory cortex. The second study delves into the dynamics of temporal processing of speech in automatic speech recognition networks, elucidating how these systems learn to integrate information across various timescales, mirroring certain aspects of biological temporal processing. The third study presents a rigorous examination of the neural encoding of linguistic information of speech in the auditory cortex during speech comprehension. By analyzing neural responses to natural speech, we identify explicit, distributed neural encoding across multiple levels of linguistic processing, from phonetic features to semantic meaning. This multilevel linguistic analysis contributes to our understanding of the hierarchical and distributed nature of speech processing in the human brain. The final chapter of this dissertation compares linguistic encoding between an automatic speech recognition system and the human brain, elucidating their computational and representational similarities and differences. This comparison underscores the nuanced understanding of how linguistic information is processed and encoded across different systems, offering insights into both biological perception and artificial intelligence mechanisms in speech processing. Through this comprehensive examination, the dissertation advances our understanding of the computational and representational foundations of speech perception, demonstrating the potential of interdisciplinary approaches that bridge neuroscience and artificial intelligence to uncover the underlying mechanisms of speech processing in both artificial and biological systems.

Neural networks are making headlines every day as the tool of the future, powering artificial intelligence programs and supporting technologies never seen before. However, the training of neural networks can take days or even weeks for bigger networks, and requires the use of super computers and GPUs in academia and industry in order to achieve state of the art results. This thesis discusses employing selective measures to determine when to backpropagate and forward propagate in order to reduce training time while maintaining classification performance. This thesis tests these new algorithms on the MNIST and CASIA datasets, and achieves successful results with both algorithms on the two datasets. The selective backpropagation algorithm shows a reduction of up to 93.3% of backpropagations completed, and the selective forward propagation algorithm shows a reduction of up to 72.90% in forward propagations and backpropagations completed compared to baseline runs of always forward propagating and backpropagating. This work also discusses employing the selective backpropagation algorithm on a modified dataset with disproportional under-representation of some classes compared to others. / Master of Science / Neural Networks are some of the most commonly used and best performing tools in machine learning. However, training them to perform well is a tedious task that can take days or even weeks, since bigger networks perform better but take exponentially longer to train. What can be done to reduce training time? Imagine a student studying for a test. The student likely solves practice problems that cover the different topics that may be covered on the test. The student then evaluates which topics he/she knew well, and forgoes extensive practice and review on those in favor of focusing on topics he/she missed or was not as confident on. This thesis discusses following a similar approach in training neural networks in order to reduce their training time needed to achieve desired performance levels.

Being able to safely control autonomous heavy machinery is of uttermost importance for the conversion of traditional machines to autonomous machines. With the continuous growth of autonomous vehicles around the globe, an increasing effort has been put into certifying autonomous vehicles in terms of reliability and safety. In this thesis, we will investigate the problem with a deviation from the planned path for an autonomous hauler from Volvo Construction Equipment. The autonomous hauler has an error within the kinematic model, the feed-forward curvature-steering controller, due to a slip-effect that comes with the third wheel-axle. The deviation can especially be seen in sharp curves, where the deviation needs to be decreased in order to make the autonomous hauler more dependable and achieve an increased accuracy when following any given path. The aim of the thesis is to develop a fully functional Artificial Neural Network that has a new steering angle as output. The hypothesis for this thesis is to use an ANN to mimic the steering of a human driver, since a real driver compensates for the slipping behavior; both because the operator knows where on the road the machine is and also in the way that a human thinks many steps ahead whilst driving. This proposed ANN will have a monitor function which ensures that the steering angle command operates within its boundaries. Hence this thesis implies that it is indeed possible to ensure that the ANN performs reliably with the help of a monitor function in a simulated environment and can thus be used in dependable systems.

Deep learning has in recent years come to dominate the previously separate fields of research in machine learning, computer vision, natural language understanding and speech recognition. Despite breakthroughs in training deep networks, there remains a lack of understanding of both the optimization and structure of deep networks. The approach advocated by many researchers in the field has been to train monolithic networks with excess complexity, and strong regularization --- an approach that leaves much to desire in efficiency. Instead we propose that carefully designing networks in consideration of our prior knowledge of the task and learned representation can improve the memory and compute efficiency of state-of-the art networks, and even improve generalization --- what we propose to denote as structural priors. We present two such novel structural priors for convolutional neural networks, and evaluate them in state-of-the-art image classification CNN architectures. The first of these methods proposes to exploit our knowledge of the low-rank nature of most filters learned for natural images by structuring a deep network to learn a collection of mostly small, low-rank, filters. The second addresses the filter/channel extents of convolutional filters, by learning filters with limited channel extents. The size of these channel-wise basis filters increases with the depth of the model, giving a novel sparse connection structure that resembles a tree root. Both methods are found to improve the generalization of these architectures while also decreasing the size and increasing the efficiency of their training and test-time computation. Finally, we present work towards conditional computation in deep neural networks, moving towards a method of automatically learning structural priors in deep networks. We propose a new discriminative learning model, conditional networks, that jointly exploit the accurate representation learning capabilities of deep neural networks with the efficient conditional computation of decision trees. Conditional networks yield smaller models, and offer test-time flexibility in the trade-off of computation vs. accuracy.

Deep neural networks with millions of parameters are at the heart of many state of the art computer vision models. However, recent works have shown that models with much smaller number of parameters can often perform just as well. A smaller model has the advantage of being faster to evaluate and easier to store - both of which are crucial for real-time and embedded applications. While prior work on compressing neural networks have looked at methods based on sparsity, quantization and factorization of neural network layers, we look at the alternate approach of pruning neurons. Training Neural Networks is often described as a kind of `black magic', as successful training requires setting the right hyper-parameter values (such as the number of neurons in a layer, depth of the network, etc ). It is often not clear what these values should be, and these decisions often end up being either ad-hoc or driven through extensive experimentation. It would be desirable to automatically set some of these hyper-parameters for the user so as to minimize trial-and-error. Combining this objective with our earlier preference for smaller models, we ask the following question - for a given task, is it possible to come up with small neural network architectures automatically? In this thesis, we propose methods to achieve the same. The work is divided into four parts. First, given a neural network, we look at the problem of identifying important and unimportant neurons. We look at this problem in a data-free setting, i.e; assuming that the data the neural network was trained on, is not available. We propose two rules for identifying wasteful neurons and show that these suffice in such a data-free setting. By removing neurons based on these rules, we are able to reduce model size without significantly affecting accuracy. Second, we propose an automated learning procedure to remove neurons during the process of training. We call this procedure ‘Architecture-Learning’, as this automatically discovers the optimal width and depth of neural networks. We empirically show that this procedure is preferable to trial-and-error based Bayesian Optimization procedures for selecting neural network architectures. Third, we connect ‘Architecture-Learning’ to a popular regularize called ‘Dropout’, and propose a novel regularized which we call ‘Generalized Dropout’. From a Bayesian viewpoint, this method corresponds to a hierarchical extension of the Dropout algorithm. Empirically, we observe that Generalized Dropout corresponds to a more flexible version of Dropout, and works in scenarios where Dropout fails. Finally, we apply our procedure for removing neurons to the problem of removing weights in a neural network, and achieve state-of-the-art results in scarifying neural networks.

<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>

The aim of this Master's thesis is to design methods of active learning and to experiment with datasets of historical documents. A large and diverse dataset IMPACT of more than one million lines is used for experiments. I am using neural networks to check the readability of lines and correctness of their annotations. Firstly, I compare architectures of convolutional and recurrent neural networks with bidirectional LSTM layer. Next, I study different ways of learning neural networks using methods of active learning. Mainly I use active learning to adapt neural networks to documents that the neural networks do not have in the original training dataset. Active learning is thus used for picking appropriate adaptation data. Convolutional neural networks achieve 98.6\% accuracy, recurrent neural networks achieve 99.5\% accuracy. Active learning decreases error by 26\% compared to random pick of adaptations data.

The aim of this work is to increase the accuracy of the transcription of text documents. This work is mainly focused on texts printed on degraded materials such as newspapers or old books. To solve this problem, the current method and problems associated with text recognition are analyzed. Based on the acquired knowledge, the implemented method based on GAN network architecture is chosen. Experiments are a performer on these networks in order to find their appropriate size and their learning parameters. Subsequently, testing is performed to compare different learning methods and compare their results. Both training and testing is a performer on an artificial data set. Using implemented trained networks increases the transcription accuracy from 65.61 % for the raw damaged text lines to 93.23 % for lines processed by this network.

První část této diplomové práci se zabývá teoretickým rozborem principů neuronových sítí, včetně možnosti jejich použití v oblasti rozpoznávání řeči. Práce pokračuje popisem viceúkolových neuronových sítí a souvisejících experimentů. Praktická část práce obsahovala změny software pro trénování neuronových sítí, které umožnily viceúkolové trénování. Je rovněž popsáno připravené prostředí, včetně několika dedikovaných skriptů. Experimenty představené v této diplomové práci ověřují použití artikulačních characteristik řeči pro viceúkolové trénování. Experimenty byly provedeny na dvou řečových databázích lišících se kvalitou a velikostí a representujících různé jazyky - angličtinu a vietnamštinu. Artikulační charakteristiky byly také kombinovány s jinými sekundárními úkoly, například kontextem, s záměrem ověřit jejich komplementaritu. Porovnaní je provedeno s neuronovými sítěmi různých velikostí tak, aby byl popsán vztah mezi velikostí neuronových sítí a efektivitou viceúkolového trénování. Závěrem provedených experimentů je, že viceúkolové trénování s použitím artikulačnich charakteristik jako sekundárních úkolů vede k lepšímu trénování neuronových sítí a výsledkem tohoto trénování může být přesnější rozpoznávání fonémů. V závěru práce jsou viceúkolové neuronové sítě testovány v systému rozpoznávání řeči jako extraktor příznaků.