UCLA

Based on their connectivity, network glasses can be classified as flexible, stressed–rigid, or isostatic, if the number of topological constraints is lower, larger, or equal to the number of atomic degrees of freedom, respectively. Thanks to the absence of any stoichiometric requirement, the rigidity of glasses can be continuously tuned (e.g., from flexible to stressed–rigid) by changing their chemical composition. Interestingly, optimally-constrained isostatic glasses have been noted to exhibit unusual properties (e.g., nearly-reversible glass transition, low relaxation, desirable mechanical properties, etc.). Especially, the non-aging intermediate state features an almost vanished endotherm between the first and second heat scan across glass transition, providing a pathway for phase change material optimization in the application of non-volatile rewriteable media. However, the physical origin of the unusual behaviors and properties of isostatic glasses remain unclear.This thesis begins with investigating how the flexible-to-rigid transition in network glasses is encoded in their energy landscape based on molecular dynamics simulations. To this end, we introduce a simplified, yet realistic model of network glasses with varying connectivity. We characterize the topography of these glasses by adopting the activation-relaxation technique (ART), which enables a systematic search of saddle points and transition pathways in the energy landscape surface. We then demonstrate that the flexible-to-rigid transition arises from an interplay between low-energy saddle points (in flexible glasses) and topological frustration (in stressed–rigid glasses). Also, by utilizing the ring structure, we expand the transition correlation with ring size distribution. Meanwhile, we highlight the local heterogeneity with all the energy landscape features by dicing the model into small cubes. Comparing within a single glass helps exclude the effect of different configurations, further consolidating our conclusion on the physical origin of rigidity transition. Finally, to explore the role of chemistry effect in rigidity transition, we compare the behavior of the simple connectivity model with a realistic GexSe1-x model. With the similar shape of enthalpic differences, the realistic model could reveal the effect of glass-forming ability with experimental results where the simple model fails. Overall, we have a clear pathway towards understanding the physical origin of rigidity transition of GexSe1-x glass.

Pediatric medical visits represent a unique opportunity for studying uncertainty, authority, and agency. In these visits medical authority and parental authority converge on a common goal — the child’s best interests. However, physicians and parents do not always agree on what courses of action are best. Physicians may disagree with parents but nevertheless rely on them to carry out treatment plans. Parents may challenge medical authority but nevertheless rely on physicians for access to the medical goods and services that they need to care for their child. In these points of departure medical authority and parental authority collide; and when the child’s problem is non-routine like a seizure the stakes can be particularly high. This dissertation explores the physician-parent partnership in a particular context: pediatric neurology visits for overnight vEEG testing. I adopt a conversation analytic approach to examine interactions between physicians and parents during these encounters, paying particular attention to the themes of uncertainty, authority, and agency. I find that parents and physicians use (un)certainty to accomplish specific interactional goals. Parents can invoke uncertainty as an account for their conduct when they have somehow challenged medical authority, thus mitigating damage to the physician-family partnership; and physicians can modulate the certainty of diagnoses, treatment efficacy, and other aspects of the child’s condition and care as a means of exerting control over visit outcomes. In the context of news deliveries, I find that the relative rights to ascribe valence to news in pediatric neurology diverge from those observed in everyday life, and this causes problems in the delivery and reception of good news. In these encounters, physicians prioritize conveying the facts of the news over characterizing its valence, but parents tend to treat both components as necessary before they are willing to assess the news. When physicians fail to provide either component, parents orient to news deliveries as incomplete. This not only causes difficulties in parents’ reception of the news but also leads to protracted news deliveries. Taken together, these findings suggest an enduring orientation to medical authority as a legitimate property of the physician-family partnership.

Nowadays, abundant data across various domains necessitate high-performance computing capabilities. While we used to be able to answer this need by scaling the frequency, the breakdown of Dennard's scaling has rendered this approach obsolete. On the other hand, Domain-specific Accelerators (DSAs) have gained a growing interest since they can offer high performance while being energy efficient. This stems from several factors, such as,1) they support utilizing special data types and operations, 2) they offer massive parallelism, 3) one can customize the memory access, 4) customizing the control/data path helps with amortizing the overhead of fixed instructions, and 5) one has the option of co-designing the algorithm with the hardware.

Unfortunately, despite the huge speedups that DSAs can deliver compared to general-purpose processors, their programmability has not caught up. In the past few decades, High-Level Synthesis (HLS) tools were introduced to raise the abstraction level and free designers from delving into architecture details at the circuit level. While HLS can significantly reduce the efforts involved in the hardware architecture design, not every HLS code yields optimal performance, requiring designers to articulate the most suitable microarchitecture for the target application. This can affect the design turnaround times as there are more choices to explore at a higher level. Moreover, this limitation has confined the DSA community primarily to hardware designers, impeding widespread adoption. This dissertation endeavors to alleviate this problem by combining customized computing and machine learning. Consequently, this dissertation consists of two core parts: 1) customized computing tailored for machine learning applications, and 2) machine learning employed to automate the optimization process of customized computing. Our focus will be on FPGAs as their cost-effective nature and rapid prototyping capabilities make them especially suitable for our research.

The large amounts of data available in data centers have motivated researchers to develop machine learning algorithms for processing them. Given that a significant portion of data stored in these centers exists in the form of images or graphs, our attention is directed towards two prominent algorithms designed for such tasks: Convolutional Neural Network (CNN) and Graph Convolutional Network (GCN). In the first part of the dissertation, we develop architecture templates for accelerating these applications. This approach facilitates a reduction in the development cycle, allowing the instantiation of module templates with customizable parameters based on the specific target application.

In the second part of the dissertation, we move our focus to general applications and work on automating their optimization steps including design space exploration and performance/area modeling. Therefore, we structure our problem in a way that can be fed into the learning algorithms. We develop a highly efficient bottleneck optimizer to explore the search space. We also explore different learning algorithms including multi-layer perceptron, graph neural networks, attention networks, jumping knowledge networks, etc., aiming to create a performance predictor that is both highly accurate and robust. Our studies show that we can optimize the microarchitecture of general applications quickly using our automated tools. This can open new doors to those without hardware knowledge to try customized computing which in turn helps to broaden the FPGA community and further improve its technology.

Post translational modifications of proteins alter the biological landscape creating functional diversity. One modification, arginine methylation, was first identified in 1968 from calf thymus hydrolysates producing guanidino-methylated arginine derivatives. However, the enzymes that produce these modifications were poorly characterized until 1996 when the genes of the first protein arginine methyltransferases were cloned from yeast and mammalian cells. At this time, a family of nine mammalian genes has been identified that encode protein arginine methyltransferases (PRMTs). In vitro experiments identified three distinct types. Type I PRMTs catalyze asymmetric dimethylarginine (ADMA) (PRMTs 1-4, 6 and 8), Type II PRMTs catalyze symmetric dimethylarginine (SDMA) (PRMT5 and 9), and the only type III PRMT that catalyzes monomethylarginine (MMA) (PRMT7). The active sites of each of the major enzymes that form ADMA, SDMA and MMA have distinct structural architectures allowing for their specificity.

In this dissertation I have focused my work on the major type I enzyme, PRMT1, the major type II enzyme, PRMT5, and the type III enzyme, PRMT7. I showed that each of these human enzymes behave differently under physiological stress conditions associated with temperature, pH, and ionic strength thus potentially leading to alterations in the proteomic arginine methylation landscape. In particular, PRMT7 is maximally active at sub-physiological temperatures and at nonphysiological pH and ionic strength, suggesting regulatory roles. I then characterized the unusual substrate specificity of the PRMT7 enzyme with peptide substrates to demonstrate the exquisite dependence upon variations of the Arg-X-Arg motif.With the identification of a PRMT7 motif in the human Fhod1 and Fhod3 actin binding proteins, I characterized methylation reactions that were dependent upon the phosphorylation state of an adjacent serine residue. These results pointed to the cross-talk that can occur between phosphorylation and methylation reactions. Interestingly, I found little or no effect of methylation on ROCK1 protein kinase activity.

PRMT enzymes have been identified to be oncogenic and closely associated with cancer progression. Surprisingly, it was found that methionine-dependent malignant cancer cells had no detectable alteration of protein arginine methylation than methionine-independent less malignant cells, suggesting that the methionine effect maybe be regulated through alternative pathways.

As the solar wind expands into the interplanetary medium, its turbulent nature changes dramatically. The synergy of the Parker Solar Probe, Solar Orbiter, and WIND missions is enabling hitherto impossible studies of plasma turbulence throughout the inner heliosphere, ranging from within the Alfv

Human real-time communication creates a limitation on the flow of information, which requires the transfer of carefully chosen and concise data in various situations. Although pointing is sparse, overloaded, and indirect, it allows humans to effectively decode shared information, (ex)change their minds, and plan accordingly. I introduce a model that explains how humans choose information for communication and understand communication by utilizing the linguistics concept of ``relevance'' derived from decision-making theory and theory of mind.

The modeling approach taken in this dissertation is inspired by many seemingly separated domains. First, I apply theory of mind from cognitive science and partially observable Markov decision process to formally model the components of human mind and how they make decisions, building a scaffold for modeling human communication. Second, I derive how humans coordinate and share their mind by applying the concepts of paternalistic helping in developmental psychology and philosophical discussion about empathy. Third, I derived the definition of utility-based relevance as how much a signaler's belief can make a positive difference to its receiver's well-being, utilizing the cooperative assumption of human communication in linguistics and comparative psychology. I conducted simulation and human behavioral experiments to show that relevance-based communication model can model the overloaded and indirect human communication and can predict humans' choices of signals in communication. Artificial intelligence agents that communicate with relevance-based models are more well-received by humans. Finally, I use Markov decision process and partially observable Markov decision process to propose a way of finding the best timing for sparse human communication.

With globalization escalating, multinational companies frequently hold meetings involving both domestic and international employees. However, time zone differences often result in international employees missing some meetings. This thesis explores an innovative solution to address this issue and ensure that colleagues who miss meetings can quickly catch up on the content. The core of this solution involves fine-tuning the Whisper model to convert audio recordings of meetings to text, followed by advanced summary transformers based on fine-tuning Llama3 and specific prompts to summarize the converted text. The resulting summaries provide a concise and comprehensive overview of the meeting's content, which can then be distributed to employees who could not attend due to time zone constraints. This approach not only enhances the efficiency of work communication among colleagues but also optimizes the global management and operational efficiency of the company.

This study investigated coffeemaker search interest in the United States using the monthlytime series data from Google Trends. The forecasting model developed can be utilized as a part of the coffeemaker market research since accurately forecasting user interest would enable whoever is intrigued to anticipate future developments and make informed decisions. To analyze the underlying pattern, the data was decomposed with STL into seasonal, trend, and residual components. We observed a consistent annual seasonality with a surge in interest every November and December. This pattern was attributed to the increase in user interest during the end-of-the-year holiday season sales. Anomaly detection using the STL residuals found two anomalies. The anomaly witnessed in December 2020 is best understood as the result of the demand surge during the holiday season compounded by the adoption of online shopping imposed by the COVID-19 lockdown. For the model selection process, ACF and PACF plots were used to make the initial judgments on the parameters of the time series model. The first round of model selection tested potential AR and MA orders. The second round of model selection tested potential seasonal AR and MA orders. SARIMA(0, 1, 2)×(1, 0, 1)12 is the final model, chosen based on AIC and BIC scores. This model was able to capture the annual seasonal pattern and meet the stationary assumption with first-order differencing. The model has a MAPE of 4.3% and a RMSE of 3.841 with the rolling forecast origin prediction on the out-sample set. The residuals were confirmed to be white noise, which indicates the SARIMA model is a good fit for predicting the monthly coffeemaker search interest in the United States.

Rewards can enhance memory for important information; however, intrinsic motivation is also an important component of long-term learning. My dissertation explores extrinsic motivation to learn such as point values awarded on memory tasks and grades assigned in classroom settings, while considering intrinsic factors that influence learning like curiosity and interest in the material being studied. I also examined how individual differences in attention, age, and study strategies impact how learners navigate what information they should prioritize when engaging with learning materials. Value-directed Remembering (VDR; Castel et al., 2002) demonstrates the potent effects of rewards on memory for important information. Point values of varying magnitudes paired with information can motivate strategic allocation of cognitive resources that can mitigate age-related deficits in memory recall. Extrinsic rewards often accompany real-world situations to motivate better performance: grades in the classroom, bonuses in the work force, points in video games, etc. However, desired behavior and information associated with rewards are not always easy to identify in real-world contexts. Schematic support or context can make rewards more meaningful, and this may be especially true for older adults who experience age-related declines in cognitive functioning (Castel, 2005). Additionally, extrinsic incentives may not always be enough to motivate all people. Some learners may need intrinsic sources of motivation to reach a goal such as curiosity, interest, or social connection. Thus, I explored whether learners could predict the value of information using rewards and schematic support to guide them, how being able to prioritize and identify important information relates to success in classroom learning, and how prior knowledge and curiosity influence what people remember. Overall, I find evidence that both younger and older adults can benefit from extrinsic rewards paired with explicit schematic knowledge to predict important information (Chapter 2), that selectivity in study strategies can be related to success in real classroom contexts (Chapter 3), and other factors like prior knowledge, curiosity, and collaboration can benefit learning (Chapter 4). Taken together, these findings suggest that learners may decide what is important to learn and remember through various extrinsic and intrinsic factors.

Mobile devices generate an enormous amount of data traffic to satisfy their computing and communications needs. To meet these demands, mobile network operators frequently need to expand their capacity, which entails significant capital costs and increased energy consumption. Motivated by this, we seek to develop cooperative systems that will bring higher communications speeds and larger computing power to mobile devices without relying on mobile network infrastructure.

In recent years, unmanned aerial vehicle (UAV) technology has garnered interest for its potential use as a communications enabler. Swarms of UAVs can be deployed as temporary relays to meet short term but high intensity communication demands from mobile users. UAV swarms can coordinate their placement to improve the capacity on the fronthaul link between users and UAVs. Algorithms for optimal placement often rely on the knowledge of channel gain across space. Hence, we developed deep learning methods for channel gain prediction across space based on measurements collected by the UAVs and 3D maps of the environment. In line with this, we also developed methods to design UAV flying paths for optimal measurement collection such that the accuracy of channel gain prediction is maximized under constraints on the distance traveled by the UAVs. Additionally, we develop a reinforcement-learning based approach that controls a UAV to directly improve the fronthaul link without relying on channel gain knowledge across space.

With the proliferation of intelligent vehicles, there is an increasing number of computationally demanding computer applications appearing in vehicular environments. Providing the computational resources to meet the demands of such applications is a critical problem. In this work, we consider a cooperative computing paradigm between intelligent vehicles of similar computing power to enable emerging vehicular applications. Vehicles cooperate with each other over vehicle-to-vehicle networks to form vehicular micro clouds that can complete computationally intensive tasks without relying on cloud or edge computing. We developed optimized resource assignment and scheduling algorithms that efficiently use vehicular computing resources for computation in emerging vehicular applications. Our proposed approaches adapt to link quality changes between vehicles and prevent congestion in vehicular networks, even in the presence of incumbent interference.