Classical value estimation reinforcement learning algorithms do not perform very well in dynamic environments. On the other hand, the reinforcement learning of animals is quite flexible: they can adapt to dynamic environments very quickly and deal with noisy inputs very effectively. One feature that may contribute to animals' good performance in dynamic environments is that they learn and perceive the time to reward. In this research, we attempt to learn and perceive the time to reward and explore situations where the learned time information can be used to improve the performance of the learning agent in dynamic environments. The type of dynamic environments that we are interested in is that type of switching environment which stays the same for a long time, then changes abruptly, and then holds for a long time before another change. The type of dynamics that we mainly focus on is the time to reward, though we also extend the ideas to learning and perceiving other criteria of optimality, e.g. the discounted return, so that they can still work even when the amount of reward may also change. Specifically, both the mean and variance of the time to reward are learned and then used to detect changes in the environment and to decide whether the agent should give up a suboptimal action. When a change in the environment is detected, the learning agent responds specifically to the change in order to recover quickly from it. When it is found that the current action is still worse than the optimal one, the agent gives up this time's exploration of the action and then remakes its decision in order to avoid longer than necessary exploration. The results of our experiments using two real-world problems show that they have effectively sped up learning, reduced the time taken to recover from environmental changes, and improved the performance of the agent after the learning converges in most of the test cases compared with classical value estimation reinforcement learning algorithms. In addition, we have successfully used spiking neurons to implement various phenomena of classical conditioning, the simplest form of animal reinforcement learning in dynamic environments, and also pointed out a possible implementation of instrumental conditioning and general reinforcement learning using similar models.

A human brain has billions of processing elements operating simultaneously; the only practical way to model this computationally is with a massively-parallel computer. A computer on such a significant scale requires hundreds of thousands of interconnected processing elements, a complex environment which requires many levels of monitoring, management and control. Management begins from the moment power is applied and continues whilst the application software loads, executes, and the results are downloaded. This is the story of the research and development of a framework of scalable management tools that support SpiNNaker, a novel computing architecture designed to model spiking neural networks of biologically-significant sizes. This management framework provides solutions from the most fundamental set of power-on self-tests, through to complex, real-time monitoring of the health of the hardware and the software during simulation. The framework devised uses standard tools where appropriate, covering hardware up / down events and capacity information, through to bespoke software developed to provide real-time insight to neural network software operation across multiple levels of abstraction. With this layered management approach, users (or automated agents) have access to results dynamically and are able to make informed decisions on required actions in real-time.

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>

Spiking neural networks (SNN) are dynamic models of biological neurons, that communicates with event-based signals called spikes. SNN that reproduce observed properties of biological senses like vision are developed to better understand how such systems function, and to learn how more efficient sensor systems can be engineered. A branching parameter describes the average probability for spikes to propagate between two different neuron populations. The adaptation of branching parameters towards critical values is known to be important for maximizing the sensitivity and dynamic range of SNN. In this thesis, a recently proposed SNN model for visual feature learning and pattern recognition known as the E-I Net model is studied and extended with a critical branching mechanism. The resulting modified E-I Net model is studied with numerical experiments and two different types of sensory queues. The experiments show that the modified E-I Net model demonstrates critical branching and power-law scaling behavior, as expected from SNN near criticality, but the power-laws are broken and the stimuli reconstruction error is higher compared to the error of the original E-I Net model. Thus, on the basis of these experiments, it is not clear how to properly extend the E-I Net model properly with a critical branching mechanism. The E-I Net model has a particular structure where the inhibitory neurons (I) are tuned to decorrelate the excitatory neurons (E) so that the visual features learned matches the angular and frequency distributions of feature detectors in visual cortex V1 and different stimuli are represented by sparse subsets of the neurons. The broken power-laws correspond to different scaling behavior at low and high spike rates, which may be related to the efficacy of inhibition in the model.

No description available.

<p>Spiking Neural Networks (SNNs) have recently attracted significant research interest as the third generation of artificial neural networks that can enable low-power event-driven data analytics. We propose ANN-SNN conversion using “soft re-set” spiking neuron model, referred to as Residual Membrane Potential (RMP) spiking neuron, which retains the “resid- ual” membrane potential above threshold at the firing instants. In addition, we propose a time-based coding scheme, named Temporal-Switch-Coding (TSC), and a corresponding TSC spiking neuron model. Each input image pixel is presented using two spikes with opposite polarity and the timing between the two spiking instants is proportional to the pixel intensity. We demonstrate near loss-less ANN-SNN conversion using RMP neurons for VGG-16, ResNet-20, and ResNet-34 SNNs on challenging datasets including CIFAR-10, CIFAR-100, and ImageNet. With the help of TSC coding, it achieves 7-14.5× less inference latency, and 30-60× fewer addition operations and memory accesses per inference across datasets compared to the state of the art (SOTA) SNN models. In the second part of the thesis, we propose a new type of recurrent neural network (RNN) architecture, named Os- cillatory Fourier Neural Network (O-FNN). We demonstrate that O-FNN is mathematically equivalent to a simplified form of Discrete Fourier Transform applied onto periodical activa- tion. In particular, the computationally intensive back-propagation through time in training is eliminated, leading to faster training while achieving the SOTA inference accuracy in a diverse group of sequential tasks. For instance, applying the proposed model to sentiment analysis on IMDB review dataset reaches 89.4% test accuracy within 5 epochs, accompanied by over 35x reduction in the model size compared to Long Short-Term Memory (LSTM). The proposed novel RNN architecture is well poised for intelligent sequential processing in resource constrained hardware.</p>

Biological neuronal networks are a key object of study in the field of computational neuroscience, and recent studies have also shown their potential applicability within artificial intelligence and robotics [1]. They come in many shapes and forms, and a well known and widely studied example is the liquid state machine from 2004 [2]. In 2019, a novel and simple connectivity rule was presented with the introduction of the SpreizerNet [3]. The connectivity of the SpreizerNet is governed by a type of gradient noise known as Perlin noise, and as such the connectivity is anisotropic but correlated. The spiking activity produced in the SpreizerNet is possibly functionally relevant, e.g. for motor control or classification of input stimuli. In 2020, it was shown to be useful for motor control [4]. In this Master’s thesis, we inquire if the spiking activity of the SpreizerNet is functionally relevant in the context of stimulus representation. We investigate how input stimulus from the MNIST handwritten digits dataset is represented in the spatio-temporal activity sequences produced by the SpreizerNet, and whether this representation is sufficient for separation. Furthermore, we consider how the parameters governing the local structure of connectivity impacts representation and separation. We find that (1) the SpreizerNet separates input stimulus at the initial stage after stimulus and (2) that separation decreases with time when the activity from dissimilar inputs becomes unified. / Biologiska neurala nätverk är ett centralt studieobjekt inom beräkningsneurovetenskapen, och nyliga studier har även visat deras potentiella applicerbarhet inom artificiell intelligens och robotik [1]. De kan formuleras på många olika sätt, och ett välkänt och vida studerat exempel är liquid state machine från 2004 [2]. 2019 presenterades en ny och enkel kopplingsregel i SpreizerNätverket [3]. Kopplingarna i SpreizerNätverket styrs av en typ av gradientbrus vid namn Perlinbrus, och som sådana är de anisotropiska men korrelerade. Spikdatan som genereras av SpreizerNätverket är möjligtvis betydelsefull för funktion, till exempel för motorisk kontroll eller separation av stimuli. 2020 visade Michaelis m. fl. att spikdatan var relevant för motorisk kontroll [4]. I denna masteruppsats frågar vi oss om spikdatan är funktionellt relevant för stimulusrepresentation. Vi undersöker hur stimulus från MNIST handwritten digits -datasetet representeras i de spatiotemporella aktivitetssekvenserna som genereras i SpreizerNätverket, och huruvida denna representation är tillräcklig för separation.Vidare betraktar vi hur parametrarna som styr den lokala kopplingsstrukturen påverkar representation och separation. Vi visar att (1) SpreizerNätverket separerar stimuli i ett initialt skede efter stimuli och (2) att separationen minskar med tid när aktiviteten från olika stimuli blir enhetlig.

<div>Machine Learning (ML) workloads, being memory and compute-intensive, consume large amounts of power running on conventional computing systems, restricting their implementations to large-scale data centers. Thus, there is a need for building domain-specific hardware primitives for energy-efficient ML processing at the edge. One such approach is in-memory computing, which eliminates frequent and unnecessary data-transfers between the memory and the compute units, by directly computing the data where it is stored. Most of the chip area is consumed by on-chip SRAMs in both conventional von-Neumann systems (e.g. CPU/GPU) as well as application-specific ICs (e.g. TPU). Thus, we propose various circuit techniques to enable a range of computations such as bitwise Boolean and arithmetic computations, binary convolution operations, non-Boolean dot-product operations, lookup-table based computations, and spiking neural network implementation - all within standard SRAM memory arrays.</div><div><br></div><div>First, we propose X-SRAM, where, by using skewed sense amplifiers, bitwise Boolean operations such as NAND/NOR/XOR/IMP etc. can be enabled within 6T and 8T SRAM arrays. Moreover, exploiting the decoupled read/write ports in 8T SRAMs, we propose read-compute-store scheme where the computed data can directly be written back in the array simultaneously. </div><div><br></div><div>Second, we propose Xcel-RAM, where we show how binary convolutions can be enabled in 10T SRAM arrays for accelerating binary neural networks. We present charge sharing approach for performing XNOR operations followed by a population count (popcount) using both analog and digital techniques, highlighting the accuracy-energy tradeoff. </div><div><br></div><div>Third, we take this concept further and propose CASH-RAM, to accelerate non-Boolean operations, such as dot-products within standard 8T-SRAM arrays by utilizing the parasitic capacitances of bitlines and sourcelines. We analyze the non-idealities that arise due to analog computations and propose a self-compensation technique which reduces the effects of non-idealities, thereby reducing the errors. </div><div><br></div><div>Fourth, we propose ROM-embedded caches, RECache, using standard 8T SRAMs, useful for lookup-table (LUT) based computations. We show that just by adding an extra word-line (WL) or a source-line (SL), the same bit-cell can store a ROM bit, as well as the usual RAM bit, while maintaining the performance and area-efficiency, thereby doubling the memory density. Further we propose SPARE, an in-memory, distributed processing architecture built on RECache, for accelerating spiking neural networks (SNNs), which often require high-order polynomials and transcendental functions for solving complex neuro-synaptic models. </div><div><br></div><div>Finally, we propose IMPULSE, a 10T-SRAM compute-in-memory (CIM) macro, specifically designed for state-of-the-art SNN inference. The inherent dynamics of the neuron membrane potential in SNNs allows processing of sequential learning tasks, avoiding the complexity of recurrent neural networks. The highly-sparse spike-based computations in such spatio-temporal data can be leveraged for energy-efficiency. However, the membrane potential incurs additional memory access bottlenecks in current SNN hardware. IMPULSE triew to tackle the above challenges. It consists of a fused weight (WMEM) and membrane potential (VMEM) memory and inherently exploits sparsity in input spikes. We propose staggered data mapping and re-configurable peripherals for handling different bit-precision requirements of WMEM and VMEM, while supporting multiple neuron functionalities. The proposed macro was fabricated in 65nm CMOS technology. We demonstrate a sentiment classification task from the IMDB dataset of movie reviews and show that the SNN achieves competitive accuracy with only a fraction of trainable parameters and effective operations compared to an LSTM network.</div><div><br></div><div>These circuit explorations to embed computations in standard memory structures shows that on-chip SRAMs can do much more than just store data and can be re-purposed as on-demand accelerators for a variety of applications. </div>