The rapid advancement of Deep Neural Networks (DNNs), computing, and sensing technology has enabled many new applications, such as the self-driving vehicle, the surveillance drone, and the robotic system. Compared to conventional edge devices (e.g. cell phone or smart home devices), these emerging devices are required to deal with much more complicated and dynamic situations in real-time with bounded computation resources. However, there are several challenges, including but not limited to efficiency, real-time adaptation, model stability, and automation of architecture design.

To tackle the challenges mentioned above, model plasticity and stability are leveraged to achieve efficient and online deep learning, especially in the scenario of learning streaming data at the edge:

First, a dynamic training scheme named Continuous Growth and Pruning (CGaP) is proposed to compress the DNNs through growing important parameters and pruning unimportant ones, achieving up to 98.1% reduction in the number of parameters.

Second, this dissertation presents Progressive Segmented Training (PST), which targets catastrophic forgetting problems in continual learning through importance sampling, model segmentation, and memory-assisted balancing. PST achieves state-of-the-art accuracy with 1.5X FLOPs reduction in the complete inference path.

Third, to facilitate online learning in real applications, acquisitive learning (AL) is further proposed to emphasize both knowledge inheritance and acquisition: the majority of the knowledge is first pre-trained in the inherited model and then adapted to acquire new knowledge. The inherited model's stability is monitored by noise injection and the landscape of the loss function, while the acquisition is realized by importance sampling and model segmentation. Compared to a conventional scheme, AL reduces accuracy drop by >10X on CIFAR-100 dataset, with 5X reduction in latency per training image and 150X reduction in training FLOPs.

Finally, this dissertation presents evolutionary neural architecture search in light of model stability (ENAS-S). ENAS-S uses a novel fitness score, which addresses not only the accuracy but also the model stability, to search for an optimal inherited model for the application of continual learning. ENAS-S outperforms hand-designed DNNs when learning from a data stream at the edge.

In summary, in this dissertation, several algorithms exploiting model plasticity and model stability are presented to improve the efficiency and accuracy of deep neural networks, especially for the scenario of continual learning.

The advances of Deep Learning (DL) achieved recently have successfully demonstrated its great potential of surpassing or close to human-level performance across multiple domains. Consequently, there exists a rising demand to deploy state-of-the-art DL algorithms, e.g., Deep Neural Networks (DNN), in real-world applications to release labors from repetitive work. On the one hand, the impressive performance achieved by the DNN normally accompanies with the drawbacks of intensive memory and power usage due to enormous model size and high computation workload, which significantly hampers their deployment on the resource-limited cyber-physical systems or edge devices. Thus, the urgent demand for enhancing the inference efficiency of DNN has also great research interests across various communities. On the other hand, scientists and engineers still have insufficient knowledge about the principles of DNN which makes it mostly be treated as a black-box. Under such circumstance, DNN is like "the sword of Damocles" where its security or fault-tolerance capability is an essential concern which cannot be circumvented.

Motivated by the aforementioned concerns, this dissertation comprehensively investigates the emerging efficiency and security issues of DNNs, from both software and hardware design perspectives. From the efficiency perspective, as the foundation technique for efficient inference of target DNN, the model compression via quantization is elaborated. In order to maximize the inference performance boost, the deployment of quantized DNN on the revolutionary Computing-in-Memory based neural accelerator is presented in a cross-layer (device/circuit/system) fashion. From the security perspective, the well known adversarial attack is investigated spanning from its original input attack form (aka. Adversarial example generation) to its parameter attack variant.

With the push for integration, a slew of modern switching power management circuits are operating at higher switching frequencies in order to reduce passive filter sizes. But while these switching regulators provide power conversion at high efficiencies, their output is prone to ripples due to the inherent switching behavior. These switching regulators use linear-low dropout regulators (LDOs) downstream to provide clean supplies. Typically, these LDOs have good power supply rejection (PSR) at lower frequencies but this degrades at higher frequencies. Therefore, some residual ripple is still manifested on the output. Because of this, high power supply rejection (PSR) with a wide rejection frequency band is becoming a critical requirement in linear low-dropout regulators (LDOs) used in complex systems- on-chip (SOCs).

Typical LDOs achieve higher PSR within their loop-bandwidth; however, their supply rejection performance degrades with reduced loop-gain outside their loop- bandwidth. The LDOs with external filtering capacitors may also have spectral peaking in their PSR response, causing excess system- level supply noise. This work presents an LDO design approach, which achieves a PSR of higher than 68 dB up to 2 MHz frequency and over a wide range of loads up to 250 mA. The wide PSR bandwidth is achieved using a current-mode feedforward ripple canceller (CFFRC) amplifier which provides up to 25 dB of PSR improvement. The feedforward path gain is inherently matched to the forward gain of the LDO, not requiring calibration. The LDO has a fast load transient response with a recovery time of 6.1μs and has a quiescent current of 5.6μA. For a full load transition, the LDO achieves settling with overshoot and undershoot voltages below 27.6 mV and 36.36 mV, respectively. The LDO is designed and fabricated in a 180 nm bipolar/CMOS/DMOS (BCD) technology. The CFFRC amplifier helps to achieve low quiescent power due to its inherent current mode nature, eliminating the need for supply ripple summing amplifiers and adaptive biasing.

Deep neural networks (DNNs) have had tremendous success in a variety of

statistical learning applications due to their vast expressive power. Most

applications run DNNs on the cloud on parallelized architectures. There is a need

for for efficient DNN inference on edge with low precision hardware and analog

accelerators. To make trained models more robust for this setting, quantization and

analog compute noise are modeled as weight space perturbations to DNNs and an

information theoretic regularization scheme is used to penalize the KL-divergence

between perturbed and unperturbed models. This regularizer has similarities to

both natural gradient descent and knowledge distillation, but has the advantage of

explicitly promoting the network to and a broader minimum that is robust to

weight space perturbations. In addition to the proposed regularization,

KL-divergence is directly minimized using knowledge distillation. Initial validation

on FashionMNIST and CIFAR10 shows that the information theoretic regularizer

and knowledge distillation outperform existing quantization schemes based on the

straight through estimator or L2 constrained quantization.

With the exponential growth in video content over the period of the last few years, analysis of videos is becoming more crucial for many applications such as self-driving cars, healthcare, and traffic management. Most of these video analysis application uses deep learning algorithms such as convolution neural networks (CNN) because of their high accuracy in object detection. Thus enhancing the performance of CNN models become crucial for video analysis. CNN models are computationally-expensive operations and often require high-end graphics processing units (GPUs) for acceleration. However, for real-time applications in an energy-thermal constrained environment such as traffic management, GPUs are less preferred because of their high power consumption, limited energy efficiency. They are challenging to fit in a small place.

To enable real-time video analytics in emerging large scale Internet of things (IoT) applications, the computation must happen at the network edge (near the cameras) in a distributed fashion. Thus, edge computing must be adopted. Recent studies have shown that field-programmable gate arrays (FPGAs) are highly suitable for edge computing due to their architecture adaptiveness, high computational throughput for streaming processing, and high energy efficiency.

This thesis presents a generic OpenCL-defined CNN accelerator architecture optimized for FPGA-based real-time video analytics on edge. The proposed CNN OpenCL kernel adopts a highly pipelined and parallelized 1-D systolic array architecture, which explores both spatial and temporal parallelism for energy efficiency CNN acceleration on FPGAs. The large fan-in and fan-out of computational units to the memory interface are identified as the limiting factor in existing designs that causes scalability issues, and solutions are proposed to resolve the issue with compiler automation. The proposed CNN kernel is highly scalable and parameterized by three architecture parameters, namely pe_num, reuse_fac, and vec_fac, which can be adapted to achieve 100% utilization of the coarse-grained computation resources (e.g., DSP blocks) for a given FPGA. The proposed CNN kernel is generic and can be used to accelerate a wide range of CNN models without recompiling the FPGA kernel hardware. The performance of Alexnet, Resnet-50, Retinanet, and Light-weight Retinanet has been measured by the proposed CNN kernel on Intel Arria 10 GX1150 FPGA. The measurement result shows that the proposed CNN kernel, when mapped with 100% utilization of computation resources, can achieve a latency of 11ms, 84ms, 1614.9ms, and 990.34ms for Alexnet, Resnet-50, Retinanet, and Light-weight Retinanet respectively when the input feature maps and weights are represented using 32-bit floating-point data type.

While machine/deep learning algorithms have been successfully used in many practical applications including object detection and image/video classification, accurate, fast, and low-power hardware implementations of such algorithms are still a challenging task, especially for mobile systems such as Internet of Things, autonomous vehicles, and smart drones.

This work presents an energy-efficient programmable application-specific integrated circuit (ASIC) accelerator for object detection. The proposed ASIC supports multi-class (face/traffic sign/car license plate/pedestrian), many-object (up to 50) in one image with different sizes (6 down-/11 up-scaling), and high accuracy (87% for face detection datasets). The proposed accelerator is composed of an integral channel detector with 2,000 classifiers for five rigid boosted templates to make a strong object detection. By jointly optimizing the algorithm and efficient hardware architecture, the prototype chip implemented in 65nm demonstrates real-time object detection of 20-50 frames/s with 22.5-181.7mW (0.54-1.75nJ/pixel) at 0.58-1.1V supply.



In this work, to reduce computation without accuracy degradation, an energy-efficient deep convolutional neural network (DCNN) accelerator is proposed based on a novel conditional computing scheme and integrates convolution with subsequent max-pooling operations. This way, the total number of bit-wise convolutions could be reduced by ~2x, without affecting the output feature values. This work also has been developing an optimized dataflow that exploits sparsity, maximizes data re-use and minimizes off-chip memory access, which can improve upon existing hardware works. The total off-chip memory access can be saved by 2.12x. Preliminary results of the proposed DCNN accelerator achieved a peak 7.35 TOPS/W for VGG-16 by post-layout simulation results in 40nm.

A number of recent efforts have attempted to design custom inference engine based on various approaches, including the systolic architecture, near memory processing, and in-meomry computing concept. This work evaluates a comprehensive comparison of these various approaches in a unified framework. This work also presents the proposed energy-efficient in-memory computing accelerator for deep neural networks (DNNs) by integrating many instances of in-memory computing macros with an ensemble of peripheral digital circuits, which supports configurable multibit activations and large-scale DNNs seamlessly while substantially improving the chip-level energy-efficiency. Proposed accelerator is fully designed in 65nm, demonstrating ultralow energy consumption for DNNs.

The past decade has seen a tremendous surge in running machine learning (ML) functions on mobile devices, from mere novelty applications to now indispensable features for the next generation of devices.

While the mobile platform capabilities range widely, long battery life and reliability are common design concerns that are crucial to remain competitive.

Consequently, state-of-the-art mobile platforms have become highly heterogeneous by combining a powerful CPUs with GPUs to accelerate the computation of deep neural networks (DNNs), which are the most common structures to perform ML operations.

But traditional von Neumann architectures are not optimized for the high memory bandwidth and massively parallel computation demands required by DNNs.

Hence, propelling research into non-von Neumann architectures to support the demands of DNNs.

The re-imagining of computer architectures to perform efficient DNN computations requires focusing on the prohibitive demands presented by DNNs and alleviating them. The two central challenges for efficient computation are (1) large memory storage and movement due to weights of the DNN and (2) massively parallel multiplications to compute the DNN output.

Introducing sparsity into the DNNs, where certain percentage of either the weights or the outputs of the DNN are zero, greatly helps with both challenges. This along with algorithm-hardware co-design to compress the DNNs is demonstrated to provide efficient solutions to greatly reduce the power consumption of hardware that compute DNNs. Additionally, exploring emerging technologies such as non-volatile memories and 3-D stacking of silicon in conjunction with algorithm-hardware co-design architectures will pave the way for the next generation of mobile devices.

Towards the objectives stated above, our specific contributions include (a) an architecture based on resistive crosspoint array that can update all values stored and compute matrix vector multiplication in parallel within a single cycle, (b) a framework of training DNNs with a block-wise sparsity to drastically reduce memory storage and total number of computations required to compute the output of DNNs, (c) the exploration of hardware implementations of sparse DNNs and architectural guidelines to reduce power consumption for the implementations in monolithic 3D integrated circuits, and (d) a prototype chip in 65nm CMOS accelerator for long-short term memory networks trained with the proposed block-wise sparsity scheme.

Deep learning (DL) has proved itself be one of the most important developements till date with far reaching impacts in numerous fields like robotics, computer vision, surveillance, speech processing, machine translation, finance, etc. They are now widely used for countless applications because of their ability to generalize real world data, robustness to noise in previously unseen data and high inference accuracy. With the ability to learn useful features from raw sensor data, deep learning algorithms have out-performed tradinal AI algorithms and pushed the boundaries of what can be achieved with AI. In this work, we demonstrate the power of deep learning by developing a neural network to automatically detect cough instances from audio recorded in un-constrained environments. For this, 24 hours long recordings from 9 dierent patients is collected and carefully labeled by medical personel. A pre-processing algorithm is proposed to convert event based cough dataset to a more informative dataset with start and end of coughs and also introduce data augmentation for regularizing the training procedure. The proposed neural network achieves 92.3% leave-one-out accuracy on data captured in real world.

Deep neural networks are composed of multiple layers that are compute/memory intensive. This makes it difficult to execute these algorithms real-time with low power consumption using existing general purpose computers. In this work, we propose hardware accelerators for a traditional AI algorithm based on random forest trees and two representative deep convolutional neural networks (AlexNet and VGG). With the proposed acceleration techniques, ~ 30x performance improvement was achieved compared to CPU for random forest trees. For deep CNNS, we demonstrate that much higher performance can be achieved with architecture space exploration using any optimization algorithms with system level performance and area models for hardware primitives as inputs and goal of minimizing latency with given resource constraints. With this method, ~30GOPs performance was achieved for Stratix V FPGA boards.

Hardware acceleration of DL algorithms alone is not always the most ecient way and sucient to achieve desired performance. There is a huge headroom available for performance improvement provided the algorithms are designed keeping in mind the hardware limitations and bottlenecks. This work achieves hardware-software co-optimization for Non-Maximal Suppression (NMS) algorithm. Using the proposed algorithmic changes and hardware architecture

With CMOS scaling coming to an end and increasing memory bandwidth bottlenecks, CMOS based system might not scale enough to accommodate requirements of more complicated and deeper neural networks in future. In this work, we explore RRAM crossbars and arrays as compact, high performing and energy efficient alternative to CMOS accelerators for deep learning training and inference. We propose and implement RRAM periphery read and write circuits and achieved ~3000x performance improvement in online dictionary learning compared to CPU.

This work also examines the realistic RRAM devices and their non-idealities. We do an in-depth study of the effects of RRAM non-idealities on inference accuracy when a pretrained model is mapped to RRAM based accelerators. To mitigate this issue, we propose Random Sparse Adaptation (RSA), a novel scheme aimed at tuning the model to take care of the faults of the RRAM array on which it is mapped. Our proposed method can achieve inference accuracy much higher than what traditional Read-Verify-Write (R-V-W) method could achieve. RSA can also recover lost inference accuracy 100x ~ 1000x faster compared to R-V-W. Using 32-bit high precision RSA cells, we achieved ~10% higher accuracy using fautly RRAM arrays compared to what can be achieved by mapping a deep network to an 32 level RRAM array with no variations.

The rapid improvement in computation capability has made deep convolutional neural networks (CNNs) a great success in recent years on many computer vision tasks with significantly improved accuracy. During the inference phase, many applications demand low latency processing of one image with strict power consumption requirement, which reduces the efficiency of GPU and other general-purpose platform, bringing opportunities for specific acceleration hardware, e.g. FPGA, by customizing the digital circuit specific for the deep learning algorithm inference. However, deploying CNNs on portable and embedded systems is still challenging due to large data volume, intensive computation, varying algorithm structures, and frequent memory accesses. This dissertation proposes a complete design methodology and framework to accelerate the inference process of various CNN algorithms on FPGA hardware with high performance, efficiency and flexibility.

As convolution contributes most operations in CNNs, the convolution acceleration scheme significantly affects the efficiency and performance of a hardware CNN accelerator. Convolution involves multiply and accumulate (MAC) operations with four levels of loops. Without fully studying the convolution loop optimization before the hardware design phase, the resulting accelerator can hardly exploit the data reuse and manage data movement efficiently. This work overcomes these barriers by quantitatively analyzing and optimizing the design objectives (e.g. memory access) of the CNN accelerator based on multiple design variables. An efficient dataflow and hardware architecture of CNN acceleration are proposed to minimize the data communication while maximizing the resource utilization to achieve high performance.

Although great performance and efficiency can be achieved by customizing the FPGA hardware for each CNN model, significant efforts and expertise are required leading to long development time, which makes it difficult to catch up with the rapid development of CNN algorithms. In this work, we present an RTL-level CNN compiler that automatically generates customized FPGA hardware for the inference tasks of various CNNs, in order to enable high-level fast prototyping of CNNs from software to FPGA and still keep the benefits of low-level hardware optimization. First, a general-purpose library of RTL modules is developed to model different operations at each layer. The integration and dataflow of physical modules are predefined in the top-level system template and reconfigured during compilation for a given CNN algorithm. The runtime control of layer-by-layer sequential computation is managed by the proposed execution schedule so that even highly irregular and complex network topology, e.g. GoogLeNet and ResNet, can be compiled. The proposed methodology is demonstrated with various CNN algorithms, e.g. NiN, VGG, GoogLeNet and ResNet, on two different standalone FPGAs achieving state-of-the art performance.

Based on the optimized acceleration strategy, there are still a lot of design options, e.g. the degree and dimension of computation parallelism, the size of on-chip buffers, and the external memory bandwidth, which impact the utilization of computation resources and data communication efficiency, and finally affect the performance and energy consumption of the accelerator. The large design space of the accelerator makes it impractical to explore the optimal design choice during the real implementation phase. Therefore, a performance model is proposed in this work to quantitatively estimate the accelerator performance and resource utilization. By this means, the performance bottleneck and design bound can be identified and the optimal design option can be explored early in the design phase.

Hardware implementation of deep neural networks is earning significant importance nowadays. Deep neural networks are mathematical models that use learning algorithms inspired by the brain. Numerous deep learning algorithms such as multi-layer perceptrons (MLP) have demonstrated human-level recognition accuracy in image and speech classification tasks. Multiple layers of processing elements called neurons with several connections between them called synapses are used to build these networks. Hence, it involves operations that exhibit a high level of parallelism making it computationally and memory intensive. Constrained by computing resources and memory, most of the applications require a neural network which utilizes less energy. Energy efficient implementation of these computationally intense algorithms on neuromorphic hardware demands a lot of architectural optimizations. One of these optimizations would be the reduction in the network size using compression and several studies investigated compression by introducing element-wise or row-/column-/block-wise sparsity via pruning and regularization. Additionally, numerous recent works have concentrated on reducing the precision of activations and weights with some reducing to a single bit. However, combining various sparsity structures with binarized or very-low-precision (2-3 bit) neural networks have not been comprehensively explored. Output activations in these deep neural network algorithms are habitually non-binary making it difficult to exploit sparsity. On the other hand, biologically realistic models like spiking neural networks (SNN) closely mimic the operations in biological nervous systems and explore new avenues for brain-like cognitive computing. These networks deal with binary spikes, and they can exploit the input-dependent sparsity or redundancy to dynamically scale the amount of computation in turn leading to energy-efficient hardware implementation. This work discusses configurable spiking neuromorphic architecture that supports multiple hidden layers exploiting hardware reuse. It also presents design techniques for minimum-area/-energy DNN hardware with minimal degradation in accuracy. Area, performance and energy results of these DNN and SNN hardware is reported for the MNIST dataset. The Neuromorphic hardware designed for SNN algorithm in 28nm CMOS demonstrates high classification accuracy (>98% on MNIST) and low energy (51.4 - 773 (nJ) per classification). The optimized DNN hardware designed in 40nm CMOS that combines 8X structured compression and 3-bit weight precision showed 98.4% accuracy at 33 (nJ) per classification.