Deep metric learning has recently shown extremely promising results in the classical data domain, creating well-separated feature spaces. This idea was also adapted to quantum computers via Quantum Metric Learning (QMeL). QMeL consists of a 2 step process with a classical model to compress the data to fit into the limited number of qubits, then train a Parameterized Quantum Circuit (PQC) to create better separation in Hilbert Space. However, on Noisy Intermediate Scale Quantum (NISQ) devices, QMeL solutions result in high circuit width and depth, both of which limit scalability. The proposed Quantum Polar Metric Learning (QPMeL ), uses a classical model to learn the parameters of the polar form of a qubit. A shallow PQC with Ry and Rz gates is then utilized to create the state and a trainable layer of ZZ(θ)-gates to learn entanglement. The circuit also computes fidelity via a SWAP Test for the proposed Fidelity Triplet Loss function, used to train both classical and quantum components. When compared to QMeL approaches, QPMeL achieves 3X better multi-class separation, while using only 1/2 the number of gates and depth. QPMeL is shown to outperform classical networks with similar configurations, presentinga promising avenue for future research on fully classical models with quantum loss functions.

Research in building agents by employing Large Language Models (LLMs) for computer control is expanding, aiming to create agents that can efficiently automate complex or repetitive computational tasks. Prior works showcased the potential of Large Language Models (LLMs) with in-context learning (ICL). However, they suffered from limited context length and poor generalization of the underlying models, which led to poor performance in long-horizon tasks, handling multiple applications and working across multiple domains. While initial work focused on extending the coding capabilities of LLMs to work with APIs to accomplish tasks, a new body of work focused on Graphical User Interface (GUI) manipulation has shown strong success in web and mobile application automation. In this work, I introduce LUCI: Large Language Model-assisted User Control Interface, a hierarchical, modular, and efficient framework to extend the capabilities of LLMs to automate GUIs. LUCI utilizes the reasoning capabilities of LLMs to decompose tasks into sub-tasks and recursively solve them. A key innovation is the application-centric approach which creates sub-tasks by first selecting the applications needed to solve the prompt. The GUI application is decomposed into a novel compressed Information-Action-Field (IAF) representation based on the underlying syntax tree. Furthermore, LUCI follows a modular structure allowing it to be extended to new platforms without any additional training as the underlying reasoning works on my IAF representations. These innovations alongside the `ensemble of LLMs' structure allow LUCI to outperform previous supervised learning (SL), reinforcement learning (RL), and LLM approaches on Miniwob++, overcoming challenges such as limited context length, exemplar memory requirements, and human intervention for task adaptability. LUCI shows a 20% improvement over the state-of-the-art (SOTA) in GUI automation on the Mind2Web benchmark. When tested in a realistic setting with over 22 commonly used applications, LUCI achieves an 80% success rate in undertaking tasks that use a subset of these applications. I also note an over 70% success rate on unseen applications, which is a less than 5% drop as compared to the fine-tuned applications.

Data-intensive systems such as big data and large machine learning (ML) systems experience serious scalability challenges due to the ever-increasing data demand from ML and analytics applications and the resource fragmentation caused by conventional monolithic server architecture. Memory and storage disaggregation emerges as a pivotal technology to address these challenges by decoupling memory and storage resources from individual servers and managing and provisioning them to applications as a shared resource pool. This dissertation investigates several important aspects of memory and storage disaggregation and proposes novel solutions to support data-intensive applications.First, caching is a fundamental way to utilize disaggregated storage, but building a large disaggregated cache is challenging because the commonly-used fix-sized cache block allocation scheme is unable to provide good cache performance with low memory overhead for diverse cloud workloads with vastly different I/O patterns. The dissertation proposes a novel adaptive cache block allocation approach that dynamically adjusts cache block sizes based on changing I/O patterns. This approach significantly improves I/O performance while reducing memory usage, outperforming traditional fixed-size cache systems in diverse cloud workloads. Evaluation shows that it improves read latency by 20% and write latency by 9%. It also reduces the amount of I/O traffic to cloud block storage by up to 74% while achieving up to 41% memory savings with only 2 ms. Second, large ML applications such as large language model (LLM) inference are memory demanding, but to support them using disaggregated memory brings challenges to memory management since disaggregated memory has higher memory access latency compared to local memory. The dissertation proposes latency-aware memory aggregation which cautiously distributes memory accesses to minimize the latency gap between local and disaggregated memory. It also proposes NUMA-aligned tensor parallelism to further improve the computing efficiency. With these optimizations, LLM inference achieves substantial speedups. For example, first token latency improves by 61%, and end-to-end latency improves by 43% for a LLM inference task which uses a model of 66 billion parameters when the batch size is 8. Finally, to address the cost, power consumption, and volatility of DRAM, the dissertation proposes to incorporate flash memory into memory pools within the disaggregation framework. By establishing a tiered memory architecture which combines fast-tier local DRAM with slow-tier DRAM and flash memory in the memory pool and effectively migrates data based on hotness across memory tiers, this approach not only reduces expenses but also maintains the overall performance and scalability of data-intensive systems. For example, with 50% saving in memory cost, the performance degradation of training ResNet50 on ImageNet dataset is only 2.68%. Together, these contributions systematically optimize the use of memory and storage disaggregation to deliver more efficient, scalable, and cost-effective systems for supporting the data explosion in today’s and future computing systems.

Pan Tilt Traffic Cameras (PTTC) are a vital component of traffic managementsystems for monitoring/surveillance. In a real world scenario, if a vehicle is in pursuit of another vehicle or an accident has occurred at an intersection causing traffic stoppages, accurate and venerable data from PTTC is necessary to quickly localize the cars on a map for adept emergency response as more and more traffic systems are getting automated using machine learning concepts. However, the position(orientation) of the PTTC with respect to the environment is often unknown as most of them lack Inertial Measurement Units or Encoders. Current State Of the Art systems 1. Demand high performance compute and use carbon footprint heavy Deep Neural Networks(DNN), 2. Are only applicable to scenarios with appropriate lane markings or only roundabouts, 3. Demand complex mathematical computations to determine focal length and optical center first before determining the pose. A compute light approach "TIPANGLE" is presented in this work. The approach uses the concept of Siamese Neural Networks(SNN) encompassing simple mathematical functions i.e., Euclidian Distance and Contrastive Loss to achieve the objective. The effectiveness of the approach is reckoned with a thorough comparison study with alternative approaches and also by executing the approach on an embedded system i.e., Raspberry Pi 3.

With the advent of new advanced analysis tools and access to related published data, it is getting more difficult for data owners to suppress private information from published data while still providing useful information. This dual problem of providing useful, accurate information and protecting it at the same time has been challenging, especially in healthcare. The data owners lack an automated resource that provides layers of protection on a published dataset with validated statistical values for usability. Differential privacy (DP) has gained a lot of attention in the past few years as a solution to the above-mentioned dual problem. DP is defined as a statistical anonymity model that can protect the data from adversarial observation while still providing intended usage. This dissertation introduces a novel DP protection mechanism called Inexact Data Cloning (IDC), which simultaneously protects and preserves information in published data while conveying source data intent. IDC preserves the privacy of the records by converting the raw data records into clonesets. The clonesets then pass through a classifier that removes potential compromising clonesets, filtering only good inexact cloneset. The mechanism of IDC is dependent on a set of privacy protection metrics called differential privacy protection metrics (DPPM), which represents the overall protection level. IDC uses two novel performance values, differential privacy protection score (DPPS) and clone classifier selection percentage (CCSP), to estimate the privacy level of protected data. In support of using IDC as a viable data security product, a software tool chain prototype, differential privacy protection architecture (DPPA), was developed to utilize the IDC. DPPA used the engineering security mechanism of IDC. DPPA is a hub which facilitates a market for data DP security mechanisms. DPPA works by incorporating standalone IDC mechanisms and provides automation, IDC protected published datasets and statistically verified IDC dataset diagnostic report. DPPA is currently doing functional, and operational benchmark processes that quantifies the DP protection of a given published dataset. The DPPA tool was recently used to test a couple of health datasets. The test results further validate the IDC mechanism as being feasible.

5G Millimeter Wave (mmWave) technology holds great promise for Connected Autonomous Vehicles (CAVs) due to its ability to achieve data rates in the Gbps range. However, mmWave suffers high beamforming overhead and requirement of line of sight (LOS) to maintain a strong connection. For Vehicle-to-Infrastructure (V2I) scenarios, where CAVs connect to roadside units (RSUs), these drawbacks become apparent. Because vehicles are dynamic, there is a large potential for link blockages, which in turn is detrimental to the connected applications running on the vehicle, such as cooperative perception and remote driver takeover. Existing RSU selection schemes base their decisions on signal strength and vehicle trajectory alone, which is not enough to prevent the blockage of links. Most recent CAVs motion planning algorithms routinely use other vehicle's near-future plans, either by explicit communication among vehicles, or by prediction. In this thesis, I make use of this knowledge (of the other vehicle's near future path plans) to further improve the RSU association mechanism for CAVs. I solve the RSU association problem by converting it to a shortest path problem with the objective to maximize the total communication bandwidth. Evaluations of B-AWARE in simulation using Simulated Urban Mobility (SUMO) and Digital twin for self-dRiving Intelligent VEhicles (DRIVE) on 12 highway and city street scenarios with varying traffic density and RSU placements show that B-AWARE results in a 1.05x improvement of the potential datarate in the average case and 1.28x in the best case vs. the state of the art. But more impressively, B-AWARE reduces the time spent with no connection by 48% in the average case and 251% in the best case as compared to the state-of-the-art methods. This is partly a result of B-AWARE reducing almost 100% of blockage occurrences in simulation.

Among the many challenges facing circuit designers in deep sub-micron technologies, power, performance, area (PPA) and process variations are perhaps the most critical. Since existing strategies for reducing power and boosting the performance of the circuit designs have already matured to saturation, it is necessary to explore alternate unconventional strategies. This investigation focuses on using perceptrons to enhance PPA in digital circuits and starts by constructing the perceptron using a combination of complementary metal-oxide-semiconductor (CMOS) and flash technology. The use of flash enables the perceptron to have a variable delay and functionality, making them robust to process, voltage, and temperature variations. By replacing parts of an application-specific integrated circuit (ASIC) with these perceptrons, improvements of up to 30% in the area and 20% in power can be achieved without affecting performance. Furthermore, the ability to vary the delay of a perceptron enables circuit designers to fix setup and hold-time violations post-fabrication, while reprogramming the functionality enables the obfuscation of the circuits. The study extends to field-programmable gate arrays (FPGAs), showing that traditional FPGA architectures can also achieve improved PPA by replacing some Look-Up-Tables (LUTs) with perceptrons. Considering that replacing parts of traditional digital circuits provides significant improvements in PPA, a natural extension was to see whether circuits built dedicatedly using perceptrons as its compute unit would lead to improvements in energy efficiency. This was demonstrated by developing perceptron-based compute elements and constructing an architecture using these elements for Quantized Neural Network acceleration. The resulting circuit delivered up to 50 times more energy efficiency compared to a CMOS-based accelerator without using standard low-power techniques such as voltage scaling and approximate computing.

Recent advances in autonomous vehicle (AV) technologies have ensured that autonomous driving will soon be present in real-world traffic. Despite the potential of AVs, many studies have shown that traffic accidents in hybrid traffic environments (where both AVs and human-driven vehicles (HVs) are present) are inevitable because of the unpredictability of human-driven vehicles. Given that eliminating accidents is impossible, an achievable goal of designing AVs is to design them in a way so that they will not be blamed for any accident in which they are involved in. This work proposes BlaFT – a Blame-Free motion planning algorithm in hybrid Traffic. BlaFT is designed to be compatible with HVs and other AVs, and will not be blamed for accidents in a structured road environment. Also, it proves that no accidents will happen if all AVs are using the BlaFT motion planner and that when in hybrid traffic, the AV using BlaFT will be blame-free even if it is involved in a collision. The work instantiated scores of BlaFT and HV vehicles in an urban road scape loop in the 'Simulation of Urban MObility', ran the simulation for several hours, and observe that as the percentage of BlaFT vehicles increases, the traffic becomes safer. Adding BlaFT vehicles to HVs also increases the efficiency of traffic as a whole by up to 34%.

With the breakdown of Dennard scaling, computer architects can no longer rely on integrated circuit energy efficiency to scale with transistor density, and must under-clock or power-gate parts of their designs in order to fit within given power budgets. Hardware accelerators may improve energy efficiency of some compute-intensive tasks, but as more tasks are accelerated, the general-purpose portions of workloads account for a larger share of execution time while also leaving less instruction, data, or task-level parallelism to exploit. Adaptive computing systems have potential to address these challenges by modifying their behavior at runtime. Adaptation requires runtime decision-making, which can be performed both in hardware and software. While software-based decision-making is more flexible and can execute higher complexity operations compared to hardware, it also incurs a significant latency and power overhead. Hardware designs are more limited in the space of decisions they can make, but have direct access to their own internal microarchitectural states and can make faster decisions, allowing for better-informed adaptation and extracting previously unobtainable performance and security benefits. In this dissertation I study (i) the viability and trade-offs of general-purpose adaptive systems, (ii) the difficulty and complexity of making adaptation decisions, and (iii) how time spent in the observation-analysis-adaptation cycle affects adaptation benefits. I introduce techniques for (a) modeling and understanding high performance computing systems and microarchitecture, (b) enabling hardware learning and decision-making through low-latency networks, and (c) on securing hardware designs using runtime decision-making. I propose an always-awake and active learning `hardware nervous system' pervasive throughout the chip that can reason about the individual hardware module performance, energy usage, and security. I present the design and implementation of (1) a reference architecture and (2) a microarchitecture-aware static binary instrumentation tool. Finally, I provide results showing (1) that runtime adaptation is a necessary to continue improving performance on general-purpose tasks, (2) that significant performance loss and performance variation happens under the ISA-level, and is unobservable without hardware support, and (3) that hardware must possess decision-making and ‘self-awareness’ capabilities at the microarchitecture level in order to efficiently use its own faculties.

Coarse-grain reconfigurable architectures (CGRAs) have shown significant improvements as hardware accelerator whilst demanding low power. Such acceleration inherits from the nature of instruction-level parallelism and exploited by many techniques. Modulo scheduling is a popular approach to software pipelining techniques that provides an efficient heuristic to accelerations on loops, repetitive regions of an application. Existing scheduling algorithms for modulo scheduling heuristic persist on loop exiting problems that limit CGRA acceleration to only loops with known trip count and no exit statements. Another notable limitation is the early exit problem, where loops can only terminate after certain iterations as CGRA moves to kernel stage. In attempts to circumvent such obstacles, COMSAT introduces a modified modulo scheduling technique that acts as an external module and can be applied to any existing scheduling/mapping algorithms with minimal hardware changes. Experiments from MiBench and Rodinia benchmark suites have shown that COMSAT achieved an average speedup of 3x in overall benchmarks and 10x speedup in kernel regions. Without COMSAT techniques, only 25% of said loops would have been able to accelerate, reducing benchmark and kernel speedups to 1.25x and 3.63x respectively.