Search

In real-world applications, inconsistent behavior displayed by AI models before and after updates can have severe consequences, particularly in safety-critical systems like autonomous driving. The updated model may exhibit regression issues, leading to incorrect predictions that were previously accurate. While regression problems have been extensively studied in various contexts, their investigation within the realm of online domain adaptation remains limited. This poses a unique challenge due to the continuous updating of the old model with incoming data streams, potentially worsening regression problems in the source domain. In this work, we address the regression problem in online domain adaptation by mitigating regression rates in the source data while adapting to a streaming target data. Our proposed approach introduces a novel loss incorporated into the Crodobo framework. This ensures that the new model learns from the source domain in a supportive manner, avoiding the acquisition of irrelevant information and preserving privacy by removing data after successful domain adaptation. While the Crodobo framework relies on raw source data, which may contain sensitive information. Hence, it does not guarantee overall privacy protection. Consequently, we also address the privacy concerns by replacing direct access to raw data with synthetic data. Through a generative model, we force the use of synthetic data, transforming online domain adaptation into a source-free process. This guarantees the confidentiality of source data, minimizes storage requirements, and enhances the practicality of the system. Our methodologies undergoe extensive evaluation on datasets such as Visda-C, COVID-DA, and MNIST - USPS. The experimental results validate the effectiveness of our approaches.

Despite the increasing power of computers, there are still complex problems that can benefit from human abilities such as cognition, reasoning, creativity, and problem-solving. These tasks can be broken down into smaller components that humans can solve individually on crowdsourcing platforms, which is a process usually described as human computation. Similarly, citizen science, a form of human computation, invites individuals, known as citizen scientists, to actively participate in scientific research. It is also possible to incorporate gamification elements or turn the tasks into puzzles that can be solved within a video game. This thesis explores using a citizen science game, GeSort, to solve the genome sorting problem. The goal of sorting genomes is to find the smallest set of evolutionary events capable of transforming one genome into another. GeSort represents the sequences of genes of the genomes being compared as series of colored shapes. This allows players to see patterns of matches and mismatches and then use a sequence of operations such as duplications, deletions and inversions to transform one genome into the other. Through the analysis of GeSort, this thesis aims to propose different guidelines and tools to help the citizen science community build improved games and applications. We started our research by evaluating the effects of including educational content in a citizen science game. Next, we developed FORGE, a framework for organizing rewards in gamified environments. Finally, we analyzed data collected through GeSort matches played over a year. Our results suggest that educational content in a citizen science game can increase players' engagement, retention and performance. The use of FORGE can speed up the implementation of a reward system, saving development time and financial resources. Analysis of the data collected through GeSort showed, among other things, that some visual patterns could confuse players, which allowed us to formulate guide

Technical Q&A sites and Large Language Models (LLMs) are valuable resources for software developers seeking knowledge. However, the provided code snippets are often uncompilable and incomplete due to unresolved types, missing libraries, and incorrect import statements. This poses a challenge for users who wish to reuse or analyze these snippets. Existing methods either do not focus on creating compilable code or have low success rates. To address this, we propose ZS4C, a lightweight approach for Zero-shot Synthesis of compilable code For incomplete Code snippets using LLMs. ZS4C operates in two stages: first, it uses an LLM, like GPT-3.5, to identify missing import statements in a snippet; second, it collaborates with a validator (e.g., compiler) to fix compilation errors caused by incorrect imports and syntax issues. We evaluated ZS4C on the StatType-SO benchmark and a new dataset, Python-SO, which includes 539 Python snippets from Stack Overflow across the 20 most popular Python libraries. ZS4C significantly outperforms existing methods, improving the compilation rate from 63% to 95.1% compared to the state-of-the-art SnR, marking a 50.1% improvement. On average, ZS4C can infer more accurate import statements (with an F1 score of 0.98) than SnR, with an improvement of 8.5% in the F1.

In Urban Search and Rescue (USAR), hazards such as structural collapse and fire can significantly endanger robots. To mitigate these risks, it is crucial to plan task allocations that adapt to dynamic environments. Most available strategies begin with evaluating the mission in advance and formulating a static plan, which may be inflexible for unforeseen changes. This project seeks to expand our lab’s existing framework by integrating Active Risk Management (ARM) into ongoing missions. The ARM module enhances adaptability by continuously monitoring environmental hazards and initiating risk mitigation tasks. Additionally, a novel method for detecting and escaping local minima allows robots to adjust their navigation patterns, preventing immobilization. The simulation environment now features realistic fire propagation, introducing a dynamic element that rigorously tests the effectiveness of the robots’ risk management strategies. By incorporating these advancements, we aim to increase mission success rates and significantly reduce robot damage in challenging USAR scenarios.

We engaged with the problem of expectation discrepancy in human-robot interaction: a known challenge in which the expectations people form when interacting with a social robot may not align with its actual capabilities. This misalignment, an expectation discrepancy, can disappoint users and hinder interaction. While research has proposed ways to mitigate expectation discrepancy, designers lack a systematic approach to analyzing and describing expectations people form of their robot. A more rigorous theoretical framework is a necessary step towards designing robots to purposefully engineering desired expectations. We consulted theories and models from psychology and sociology on expectations between people, and conducted a survey of expectations in human-robot interactions. Through this we developed an analytical framework consisting of a novel model of the cognitive process of human-robot expectation formation, as well as a taxonomy for classifying the types of expectations they form. We finally propose preliminary methods for designers to use this framework as a tool to support systematic analysis of how and why people form expectations of a given robot and what those expectations may be. Such understanding can empower designers with greater control over people’s expectations, enabling them to combat problems of expectation discrepancy.

Deep neural networks (DNNs) are gaining popularity in a wide range of domains, ranging from speech and video recognition to healthcare. With this increased adoption comes the pressing need for securing DNN execution environments on CPUs, GPUs, and ASICs. While there are active research efforts in supporting a trusted execution environment (TEE) on CPUs, the exploration in supporting TEEs on accelerators is limited, with only a few solutions available [18, 19, 27]. A key limitation along this line of work is that these secure DNN accelerators narrowly consider a few specific architectures. The design choices and the associated cost for securing these architectures do not transfer to other diverse architectures. This paper strives to address this limitation by developing a design space exploration tool for supporting TEEs on diverse DNN accelerators. We target secure DNN accelerators equipped with cryptographic engines where the cryptographic operations are closely coupled with the data movement in the accelerators. These operations significantly complicate the scheduling for DNN accelerators, as the scheduling needs to account for the extra on-chip computation and off-chip memory accesses introduced by these cryptographic operations, and even needs to account for potential interactions across DNN layers. We tackle these challenges in our tool, called SecureLoop, by introducing a scheduling search engine with the following attributes: 1) considers the cryptographic overhead associated with every off-chip data access, 2) uses an efficient modular arithmetic technique to compute the optimal authentication block assignment for each individual layer, and 3) uses a simulated annealing algorithm to perform cross-layer optimizations. Compared to the conventional schedulers, our tool finds the schedule for secure DNN designs with up to 33.2% speedup and 50.2% improvement of energy-delay-product.

Holographic displays generate Three-Dimensional (3D) images by displaying single-color holograms time-sequentially, each lit by a single-color light source. However, representing each color one by one limits brightness in holographic displays. This paper introduces a new driving scheme for realizing brighter images in holographic displays. Unlike the conventional driving scheme, our method utilizes three light sources to illuminate each displayed hologram simultaneously at various intensity levels. In this way, our method reconstructs a multiplanar three-dimensional target scene using consecutive multi-color holograms and persistence of vision. We co-optimize multi-color holograms and required intensity levels from each light source using a gradient descent-based optimizer with a combination of application-specific loss terms. We experimentally demonstrate that our method can increase the intensity levels in holographic displays up to three times, reaching a broader range and unlocking new potentials for perceptual realism in holographic displays.

Background: Unstructured text data (UTD) in electronic medical records (EMRs) may be challenging to use because of noise including spelling errors, abbreviations, and punctuation symbols. It is expected that preprocessing improves the performance of statistical or machine learning models. Objectives: The research objectives were to assess the effect of the number and order of preprocessing methods (1) on the detection of health conditions from UTD, (2) on clinical and demographic cohort selection criteria in UTD, (3) on the similarity of information contained in pairs of EMR notes for the same patient, and (4) on accurate de-identification of UTD. Method: Study data were from the Informatics for Integrating Biology and the Bedside (i2b2) challenges. The 2008, 2014 and 2018 i2b2 datasets were used for different Objectives. Preprocessing methods included tokenization, removing punctuation symbols, correcting spelling errors, expanding abbreviations, word stemming, and lemmatization. A nested experimental design was adopted, in which order was nested within the number of methods. Balanced random forest, support vector machine, and bidirectional long short-term memory-conditional random field models were used for Objectives 1, 2, and 4, respectively, and model performance was evaluated by accuracy, sensitivity, specificity, F1 score, and precision. For Objective 3, cosine similarity was used to measure the similarity between pairs of notes. Analysis of Variance (ANOVA) and descriptive statistics were used to test research hypotheses. Results: Mean sensitivity, specificity, F1 score, accuracy, and precision were similar across the orders of methods and numbers of methods, for Objectives 1 and 2. Cosine similarity scores were similar across orders of methods for Objective 3. Deep learning models for Objective 4 were not trainable with the preprocessed data. The ANOVA F test results showed no significant effect of the order of methods for different numbers and identical me

Emergency departments are among the most crowded facilities in healthcare premises, where they receive a variety of cases, including critical ones and life-threatening conditions. Arranging the order of received patients and providing timely and efficient care is of utmost importance. This procedure is usually carried out by a nurse who considers the patient’s symptoms and vital signs besides ready resources. Existing literature revealed that there is variability in the accuracy of the triage process inside emergencies for a variety of reasons. Therefore, developing an aid tool based on Machine Learning (ML) algorithms would help mitigate this issue and improve the workflow inside such a crucial setting. This work provides a comparison between several ML-based classifying models that were developed from MIMIC-IV-ED and MIMIC-IV databases. Moreover, it presents insights into hidden patterns that explain some outcomes of subgroups in the examined individuals.

Zombies and Survivors is a two-player graph game introduced by Fitzpatrick et al. as a variant of Cops and Robbers. These games can be used to model situations involving pursuit-evasion or search-and-rescue. Two players take turns to move their pieces in a given graph, with the condition that the zombie pieces must move toward the survivor along a shortest path during the zombie player’s turn. The minimum number of zombies that is sufficient to catch the survivor is called the graph's zombie number. First, we give a construction of graphs with zombie number equal to any given k, and extend the result to a construction of graphs that simultaneously have zombie number equal to any given k and cop number equal to any given m. Then, we investigate specific graph classes. We determine the exact zombie number of the Cartesian product of a cycle and a path, and show that this quantity depends on the parity of the cycle. Upper bounds on the zombie number for strong products, Cartesian products and lexicographic products are given. We also show that the throttling numbers of Cartesian products and strong products of two paths are asymptotic to a sublinear polynomial in the length of the path. Finally, the class of graphs formed by a set of cycles with one common center vertex is investigated, and we give a characterization of such graphs that have zombie number 1 or 2.

The rapid pace of electrification and adoption of renewable energy sources necessitate the development of efficient, cost-effective, and compact power electronic converters. Among the essential components in modern converters, magnetic components play a crucial role, with their size and quantity varying depending on the application, power rating, and converter topology. Reducing the size and number of magnetic components lead to higher converter efficiency, lower cost, and higher power density. This thesis presents two distinct methods for reducing the size and quantity of magnetic components in single-phase and three-phase grid-connected converters. In single-phase converters, an innovative interleaving method is proposed, which effectively increases the power rating of a commercial Gallium Nitride (GaN)-based Power Factor Correction (PFC) converter. The proposed interleaving method along with GaN transistor and an advanced controller operating in Discontinuous Conduction Mode (DCM) lead to substantial inductor size reduction in the PFC converter and Electromagnetic Interference (EMI) filter. For three-phase converters, a novel set of topologies known as "reconfigurable filter" is introduced, minimizing the number of magnetic components. These reconfigurable filters construct LCL filters using only three inductors, a notable reduction compared to the conventional LCL filters requiring six inductors. The key components of the reconfigurable filters are a set of low-frequency bidirectional switches which lead to full utilization of existing magnetic components by changing their roles from the grid side to the converter side. The proposed topologies achieve comparable Total Harmonic Distortion (THD) performance compared to conventional LCL filters while utilizing three fewer inductors. Additionally, the reconfigurable filters suppress leakage current and offer reactive power support, making them versatile for diverse applications. The proposed topologies are applied i

Many real world optimization problems involve multiple conflicting objectives, constraints and parameters. Multi-objective optimization (MOO) techniques are used to solve these problems. The goal of MOO is to find a set of optimal solutions, or the Pareto optimal front. Multi-objective evolutionary algorithms are heuristics that evolve a population of candidate solutions to find the Pareto optimal front in a single run. The selection criterion used to select individuals in the population play an important role in determining the quality of the solutions. Pareto-based algorithms use the Pareto selection criterion to evolve different parts of the solution space introducing diverse solutions, but converge slowly to the optimal front. On the other hand, non- Pareto selection Criterion (NPC) algorithms converge faster to the Pareto front, but in the process eliminate other diverse solutions. To compensate for the strengths and weaknesses of PC and NPC, hybrid frameworks such as BCE (bi-criterion evolutionary) have been proposed. In BCE, the PC and NPC algorithms evolve separately, but also co-operate by exchanging information to explore and exploit the objective space. In the literature, two well-known evolutionary algorithms, Non-dominated Sorting Genetic Algorithm II (NSGA-II) (PC) and Multi-objective Evolutionary Algorithm based on Decomposition (MOEA/D) (NPC) have been used as a case study in the BCE framework. However, the individual algorithms are computationally expensive. In this thesis, we study the parallelization of the BCE framework. NSGA-II is highly data parallel, and is well suited for single instruction multiple data architectures. MOEA/D is non-data parallel with some parts of the algorithm being sequential. Therefore, we design the parallel NSGA-II algorithm on the GPU multi-core accelerator and parallel MOEA/D algorithm on multi-core CPU machines using an island model. Using the travelling salesperson benchmark data sets we analyze the performance of t

Dynamic environments pose many challenges as the search space is irregular, un- structured, with the data and problem space changing over time. The algorithms executing on these environments should adapt to the varying dynamic conditions. In this research we consider Ant Colony Optimization algorithm (ACO), a technique inspired by real ants in nature and therefore, should be adaptable to dynamic environ- ments. However, some studies in the literature show the contrary. Population-based ACO was introduced, a hybrid technique that combines concepts from Genetic Algo- rithms for solving problems in dynamic environments. In this thesis, we argue and show that ACO is as good as PACO or even better in some cases, by incorporating lo- cal search techniques to exploit the search space, tuning parameters in the algorithm to explore the search space and, using migration between multiple colonies (or island model) for convergence. The multiple colonies are implemented on multiple GPUs for efficiency. We perform various experiments on a dynamic travelling salesperson dataset and compare ACO and PACO with local search and island model. We also show that the parameter tuning has a significant influence on the accuracy.

Given a set S of objects, each moving with linear motion in R^d, consider the diameter D(S, t) of S at time t. In this thesis we explore optimization of extent and proximity measures of S. For instance, one possibility is to identify minimum diameter D(S, t) of S over the domain of time t. D(S, t) is an example of a measure of extent of S. On the basis of this model, other geometric measures could also be explored to be optimized for sets of objects in motion. Let n be the cardinality of S and let M(S, t) be a geometric measure of extent or proximity at time t. Given an integer k, select a subset Q ⊂ S such that |Q| = k and Q has extreme measure M(Q, t) over all possible subsets Q of cardinality k. The present thesis focuses on minimizing and maximizing M(Q, t), in one and two dimensions (d = 1 or d = 2), for which the measure corresponds to set diameter, set width, minimum axis-aligned bounding box, and minimum enclosing disk. For each measure, exact polynomial-time algorithms are proposed for selecting an optimal subset of S and finding the time t of which the subset optimizes the measure.

The world is dynamic, so are big data. The evolving challenges of managing big data volume, variety, veracity, validity, and velocity has resulted in several studies focusing on solving one or more of these perplexing issues. In this Ph.D. research, I focus on the evolving issues arising from big data variety, veracity, privacy, and accessibility. First, I design a conceptual model for capturing and storing variety of big data types including structured, semi-structured and unstructured data types and in addition, design a metadata collection framework for managing the big data in support of machine learning and open data FAIR principle of Findable, Accessibility, Interoperability and Re-usability such that the information about the data are available beyond the life cycle of the data. Second, I design hierarchical spatial-temporal model (HSTM) for managing individual record in big data in the aforementioned open data lake architecture with metadata collection framework. Third, I extend the HSTM and design the resulting hierarchical spatial-temporal privacy preserving model (HSTPPM) for preserving privacy of individual record in big data. Fourth, I extend and design applications of the HSTPPM to big data co-occurrence pattern mining and big data visualization.

This thesis explores interconnected advanced machine learning (ML) and data science (DS) methodologies for improved predictive accuracy in financial markets and resilient portfolio optimization. Studying the literature on ML/DS methodologies extensively led us to observe a significant lack of application of these advances, such as autoencoder (AE), recurrent neural networks (RNN), etc. in the finance industry. The novelty of this thesis is to study price prediction and portfolio optimization with RNN and AE algorithms. Furthermore, unsupervised ML strategies were studied to introduce robustness in portfolio optimization. For this purpose, two innovative encoder-decoder-based RNN architectures autoencoder-based gated recurrent unit (AE-GRU) and autoencoder-based long short-term memory (AE-LSTM) were proposed, which were shown to be effective in predictive efficacy across diverse asset types and market conditions, showcasing enhanced predictive accuracy for financial assets. Various DS concepts, such as data visualization, Bollinger bands, data-driven volatility estimates, unsupervised ML, etc. were integrated while implementing and experimenting with new architectures for price prediction and portfolio optimization. The proposed models in this thesis showed effectiveness in price prediction and portfolio optimization under varying market conditions. The study also highlights the benefits of diversified portfolios by proposing a novel DL-based model for portfolio construction, especially when coupled with affinity propagation (AP) clustering and appropriate data-driven risk measures based on volatility estimates - with sign correlation (VES) and volatility correlation (VEV). Traditional models optimize portfolio weights using objective functions, while recent innovations emphasize data-driven risk measures for minimum risk weights from random samples. Despite challenges with short-term data featuring negative mean returns, the proposed ML-based diversification approac

This thesis explores two unique ways of using machine learning to improve or facili- tate microwave imaging. The first contribution provides a means of calibration under the assumption of an uncooperative imaging system, a term indicating that it is im- possible or impractical to have any control over the region of interest, which means no known targets are available for calibration. The approach uses a method of style transfer provided by Cycle Generative Adversarial Networks. Cycle Generative Ad- versarial Networks are capable of learning arbitrary transformations between two sets of data in which there exists no paired samples across the two domains. In the case of calibration the two domains are synthetic field data and raw S-parameters from a physical imaging setup. The method is shown to be nearly as good as calibration with a known target for a 2D TM experimental imaging setup. The second contribution of this thesis focuses on how multi-frequency data should best be used in a machine learning model to solve the inverse problem. This work introduces two novel architectures capable of using multi-frequency data and testing the results on experimental data. The first method uses a system of sequential U- Nets referred to as the cascaded multi-frequency approach. This method turned out to be very similar to a recurrent neural network which inspired the creation of a novel LSTM-like architecture - the second method. These two methods were compared i to single frequency inversions and a ‘naive’ multi-frequency inversion which collapse the frequency data into the channels of the single frequency inversion. The models were tested on the same 2D TM experimental imaging setup in which labeled data was acquired through an automated target positioning system. Both multi-frequency approaches show significant improvement over the single frequency counter part and the naive approach.

Bitcoin, envisioned as a decentralized currency, facilitates secure micro-payments through a distributed consensus and ensures security via its widespread network. However, prioritizing security and decentralization imposes constraints on performance. My thesis explores strategies to boost the blockchain's throughput and performance alongside improving miners' financial incentives. We analyze the effects of increasing block size on throughput and security and how strategic transaction selection by miners before mining can elevate fee collection. We introduce a novel transaction-selection strategy that produces high-quality blocks more efficiently by avoiding traditional sorting and examining the benefits of regularly updating transaction sets to maximize fees. Additionally, we assess the Lightning Network's impact on reducing blockchain load and transaction costs and its potential to lower miners' fee revenues and profits. This study balances the Lightning Network's adoption and integration with third-layer applications to mitigate transaction migration from the blockchain. This thesis contributes to blockchain scalability and efficiency, presenting new transaction selection and block formation strategies that enhance network performance and miners' incentives. It also navigates the economic implications of the Lightning Network, offering a nuanced view of its effects on the blockchain ecosystem.

This thesis work presents a new approach to malware analysis by creating a specialized sandbox environment for executing and monitoring malware on a host operating system. Over 200 malware samples, along with benignware, were tested in this environment. Application Programming Interface (API) calls were traced of how these executable interacted with the host system, which including registry changes, file system access, and thread activity. Two new methods to measure complexity, Mass Radius and Radius of Gyration Fractal Dimension (FD), were developed and added to a Deep Learning model. These complexity methods helped the model converge faster. Tests showed that the Mass Radius FD measure improved convergence and accuracy over the Radius of Gyration FD in identifying malware. The complexity models performed better than standard models across different datasets. The study also found that shorter sequences of API calls and file events were more likely to indicate malicious behavior. Using GNNExplainer, linked API sequences were linked to specific malware techniques, providing deeper insights into the model’s predictions. The complexity models identified flaws in traditional methods and successfully flagged malware that other commercial sandbox methods were not able to identify, lending credence to the sophistication and applicability of this work.

The popularity of immersive web virtual worlds has increased as a result of the pandemic. Platforms like Gather, that still sees more than a million visitors quarterly post-pandemic, combined real-time video conferencing functionality with virtual worlds: multi-user virtual environments that function as virtual offices or social spaces. This combination has not only incorporated interactivity and engagement in the videoconferencing space but has also brought the capability of virtual world building to the masses, including to people without programming backgrounds. My research investigates the tools used for virtual world building, which are feature-rich creativity support tools. Given their availability, I want to understand how easy it is for non-programmers to use virtual world building tools to build virtual worlds, or more specifically, what the limitations of these virtual world building tools are by having users with backgrounds designing physical world spaces (such as architecture and interior design students) build virtual spaces. In this research, I investigate the technical challenges and creative limitations that such users experience, and how they troubleshoot their virtual world designs. The results show that participants are able to produce interesting virtual worlds with 2D virtual world authoring tools. However, advanced features appear to be too complicated to use consistently. This led to users sticking to basic features that they are familiar with. Additionally, users did not appear to test and debug their work frequently, if not at all. Based on these findings, I explain in what way these virtual world authoring tools frustrate users and produce design suggestions to improve creativity support for them. This thesis contributes results from the user study of non-programmer use of 2D virtual world authoring tools, provides evidence of user difficulty with juggling multiple toolchains and design suggestions to improve virtual world authoring experi

We introduce SLANG.D, an extension to the Slang shading language that incorporates first-class automatic differentiation support. The new shading language allows us to transform a Direct3D-based path tracer to be fully differentiable with minor modifications to existing code. SLANG.D enables a shared ecosystem between machine learning frameworks and pre-existing graphics hardware API-based rendering systems, promoting the interchange of components and ideas across these two domains. Our contributions include a differentiable type system designed to ensure type safety and semantic clarity in codebases that blend differentiable and non-differentiable code, language primitives that automatically generate both forward and reverse gradient propagation methods, and a compiler architecture that generates efficient derivative propagation shader code for graphics pipelines. Our compiler supports differentiating code that involves arbitrary control-flow, dynamic dispatch, generics and higher-order differentiation, while providing developers flexible control of checkpointing and gradient aggregation strategies for best performance. Our system allows us to differentiate an existing real-time path tracer, Falcor, with minimal changes to its shader code. We show that the compiler-generated derivative kernels perform as efficiently as handwritten ones. In several benchmarks, the SLANG.D code achieves significant speedup when compared to prior automatic differentiation systems.

2022 International Electron Devices Meeting (IEDM), San Francisco, CA, USA, III-Nitrides are attractive as field emission devices for high frequency, high power, and harsh environment applications. A wet-based digital etching and a novel device geometry was used to demonstrate GaN vertical self-alignedgate (SAG) field emitter arrays (FEA) with uniform tips of sub- 10 nm tip radius. The best GaN FEA has a current density (JA) of 10 A/cm2 at VGE = 50 V, which is better than the state-of-the-art Si field emitter arrays at the same bias condition., Department of Defense (DoD)

Patients with cancer undergoing chemotherapy frequently experience a neurological condition known as chemotherapy-related cognitive impairment, or “chemobrain,” which can persist for the remainder of their lives. Despite the growing prevalence of chemobrain, both its underlying mechanisms and treatment strategies remain poorly understood. Recent findings suggest that chemobrain shares several characteristics with neurodegenerative diseases, including chronic neuroinflammation, DNA damage, and synaptic loss. We investigated whether a noninvasive sensory stimulation treatment we term gamma entrainment using sensory stimuli (GENUS), which has been shown to alleviate aberrant immune and synaptic pathologies in mouse models of neurodegeneration, could also mitigate chemobrain phenotypes in mice administered a chemotherapeutic drug. When administered concurrently with the chemotherapeutic agent cisplatin, GENUS alleviated cisplatin-induced brain pathology, promoted oligodendrocyte survival, and improved cognitive function in a mouse model of chemobrain. These effects persisted for up to 105 days after GENUS treatment, suggesting the potential for long-lasting benefits. However, when administered to mice 90 days after chemotherapy, GENUS treatment only provided limited benefits, indicating that it was most effective when used to prevent the progression of chemobrain pathology. Furthermore, we demonstrated that the effects of GENUS in mice were not limited to cisplatin-induced chemobrain but also extended to methotrexate-induced chemobrain. Collectively, these findings suggest that GENUS may represent a versatile approach for treating chemobrain induced by different chemotherapy agents.

Linear reduced-order modeling (ROM) simplifies complex simulations by approximating the behavior of a system using a simplified kinematic representation. Typically, ROM is trained on input simulations created with a specific spatial discretization, and then serves to accelerate simulations with the same discretization. This discretization-dependence is restrictive. Becoming independent of a specific discretization would provide flexibility to mix and match mesh resolutions, connectivity, and type (tetrahedral, hexahedral) in training data; to accelerate simulations with novel discretizations unseen during training; and to accelerate adaptive simulations that temporally or parametrically change the discretization. We present a flexible, discretization-independent approach to reduced-order modeling. Like traditional ROM, we represent the configuration as a linear combination of displacement fields. Unlike traditional ROM, our displacement fields are continuous maps from every point on the reference domain to a corresponding displacement vector; these maps are represented as implicit neural fields. With linear continuous ROM (LiCROM), our training set can include multiple geometries undergoing multiple loading conditions, independent of their discretization. This opens the door to novel applications of reduced order modeling. We can now accelerate simulations that modify the geometry at runtime, for instance via cutting, hole punching, and even swapping the entire mesh. We can also accelerate simulations of geometries unseen during training. We demonstrate one-shot generalization, training on a single geometry and subsequently simulating various unseen geometries.

Sparse tensor algebra is a challenging class of workloads to accelerate due to low arithmetic intensity and varying sparsity patterns. Prior sparse tensor algebra accelerators have explored tiling sparse data to increase exploitable data reuse and improve throughput, but typically allocate tile size in a given buffer for the worst-case data occupancy. This severely limits the utilization of available memory resources and reduces data reuse. Other accelerators employ complex tiling during preprocessing or at runtime to determine the exact tile size based on its occupancy. This paper proposes a speculative tensor tiling approach, called overbooking, to improve buffer utilization by taking advantage of the distribution of nonzero elements in sparse tensors to construct larger tiles with greater data reuse. To ensure correctness, we propose a low-overhead hardware mechanism, Tailors, that can tolerate data overflow by design while ensuring reasonable data reuse. We demonstrate that Tailors can be easily integrated into the memory hierarchy of an existing sparse tensor algebra accelerator. To ensure high buffer utilization with minimal tiling overhead, we introduce a statistical approach, Swiftiles, to pick a tile size so that tiles usually fit within the buffer’s capacity, but can potentially overflow, i.e., it overbooks the buffers. Across a suite of 22 sparse tensor algebra workloads, we show that our proposed overbooking strategy introduces an average speedup of 52.7 × and 2.3 × and an average energy reduction of 22.5 × and 2.5 × over ExTensor without and with optimized tiling, respectively.

Fast simulation of digital circuits is crucial to build modern chips. But RTL (Register-Transfer-Level) simulators are slow, as they cannot exploit multicores well. Slow simulation lengthens chip design time and makes bugs more frequent. We present ASH, a parallel architecture tailored to simulation workloads. ASH consists of a tightly codesigned hardware architecture and compiler for RTL simulation. ASH exploits two key opportunities. First, it performs dataflow execution of small tasks to leverage the fine-grained parallelism in simulation workloads. Second, it performs selective event-driven execution to run only the fraction of the design exercised each cycle, skipping ineffectual tasks. ASH hardware provides a novel combination of dataflow and speculative execution, and ASH’s compiler features several novel techniques to automatically leverage this hardware. We evaluate ASH in simulation using large Verilog designs. An ASH chip with 256 simple cores is gmean 1,485 × faster than 1-core Verilator, and it is 32 × faster than parallel Verilator on a server CPU with 32 complex cores, while using 3 × less area.

Point-of-care testing using paper-based lateral flow assays (LFAs) has emerged as an attractive diagnostic platform. However, detecting small molecules such as cortisol using LFAs is challenging due to limited binding sites and weak signal generation. Here, we report the development of cortisol-specific aptamer-based probes and a paper-based origami preconcentrator (POP) to amplify the probe signal. The cortisol-specific aptamers were conjugated onto gold nanoparticles and hybridized with signal probes to create the cortisol-specific signal-on probe. POP, consisting of patterned layers with convergent wicking zones, induces electrokinetic preconcentration of the released signaling probes. By integrating cortisol-selective aptamer-based probes and POP, we accurately diagnosed cortisol levels within 30 min of signal probe incubation, followed by 10 min of preconcentration. Our sensor was able to detect cortisol levels in the range of 25–1000 ng/mL, with typical cortisol levels in plasma ranging from 40 to 250 ng/mL falling within this range. The successful detection of the wide range of cortisol samples using this approach highlights the potential of this platform as a point-of-care testing tool, particularly for lateral flow assay-based detection of small molecules like cortisol. Our approach offers a convenient and reliable method of cortisol level testing with a portable and accessible diagnosis device.

On-device learning and efficient fine-tuning enable continuous and privacy-preserving customization (e.g., locally fine-tuning large language models on personalized data). However, existing training frameworks are designed for cloud servers with powerful accelerators (e.g., GPUs, TPUs) and lack the optimizations for learning on the edge, which faces challenges of resource limitations and edge hardware diversity. We introduce PockEngine: a tiny, sparse and efficient engine to enable fine-tuning on various edge devices. PockEngine supports sparse backpropagation: it prunes the backward graph and sparsely updates the model with measured memory saving and latency reduction while maintaining the model quality. Secondly, PockEngine is compilation first: the entire training graph (including forward, backward and optimization steps) is derived at compile-time, which reduces the runtime overhead and brings opportunities for graph transformations. PockEngine also integrates a rich set of training graph optimizations, thus can further accelerate the training cost, including operator reordering and backend switching. PockEngine supports diverse applications, frontends and hardware backends: it flexibly compiles and tunes models defined in PyTorch/TensorFlow/Jax and deploys binaries to mobile CPU/GPU/DSPs. We evaluated PockEngine on both vision models and large language models. PockEngine achieves up to 15 × speedup over off-the-shelf TensorFlow (Raspberry Pi), 5.6 × memory saving back-propagation (Jetson AGX Orin). Remarkably, PockEngine enables fine-tuning LLaMav2-7B on NVIDIA Jetson AGX Orin at 550 tokens/s, 7.9 × faster than the PyTorch.

Over the past few years, the explosion in sparse tensor algebra workloads has led to a corresponding rise in domain-specific accelerators to service them. Due to the irregularity present in sparse tensors, these accelerators employ a wide variety of novel solutions to achieve good performance. At the same time, prior work on design-flexible sparse accelerator modeling does not express this full range of design features, making it difficult to understand the impact of each design choice and compare or extend the state-of-the-art.To address this, we propose TeAAL: a language and simulator generator for the concise and precise specification and evaluation of sparse tensor algebra accelerators. We use TeAAL to represent and evaluate four disparate state-of-the-art accelerators—ExTensor, Gamma, OuterSPACE, and SIGMA—and verify that it reproduces their performance with high accuracy. Finally, we demonstrate the potential of TeAAL as a tool for designing new accelerators by showing how it can be used to speed up vertex-centric programming accelerators—achieving 1.9 × on BFS and 1.2 × on SSSP over GraphDynS.

We propose a hybrid neural network and physics framework for reduced-order modeling of elastoplasticity and fracture. State-of-the-art scientific computing models like the Material Point Method (MPM) faithfully simulate large-deformation elastoplasticity and fracture mechanics. However, their long runtime and large memory consumption render them unsuitable for applications constrained by computation time and memory usage, e.g., virtual reality. To overcome these barriers, we propose a reduced-order framework. Our key innovation is training a low-dimensional manifold for the Kirchhoff stress field via an implicit neural representation. This low-dimensional neural stress field (NSF) enables efficient evaluations of stress values and, correspondingly, internal forces at arbitrary spatial locations. In addition, we also train neural deformation and affine fields to build low-dimensional manifolds for the deformation and affine momentum fields. These neural stress, deformation, and affine fields share the same low-dimensional latent space, which uniquely embeds the high-dimensional simulation state. After training, we run new simulations by evolving in this single latent space, which drastically reduces the computation time and memory consumption. Our general continuum-mechanics-based reduced-order framework is applicable to any phenomena governed by the elastodynamics equation. To showcase the versatility of our framework, we simulate a wide range of material behaviors, including elastica, sand, metal, non-Newtonian fluids, fracture, contact, and collision. We demonstrate dimension reduction by up to 100,000 × and time savings by up to 10 ×.

We report graphene films composed mostly of one or two layers of graphene grown by controlled carbon precipitation on the surface of polycrystalline Ni thin films during atmospheric chemical vapor deposition (CVD). Controlling both the methane concentration during CVD and the substrate cooling rate during graphene growth can significantly improve the thickness uniformity. As a result, one- or two- layer graphene regions occupy up to 87% of the film area. Single layer coverage accounts for 5%–11% of the overall film. These regions expand across multiple grain boundaries of the underlying polycrystalline Ni film. The number density of sites with multilayer graphene/graphite (>2 layers) is reduced as the cooling rate decreases. These films can also be transferred to other substrates and their sizes are only limited by the sizes of the Ni film and the CVD chamber. Here, we demonstrate the formation of films as large as 1 in2. These findings represent an important step towards the fabrication of large-scale high-quality graphene samples., National Science Foundation (U.S.) (CTS 05-06830), National Science Foundation (U.S.) (DMR07-04197)

2024 IEEE Custom Integrated Circuits Conference April 21st – 24th, 2024 Denver, CO U.S., Machine learning (ML) accelerators provide energy efficient neural network (NN) implementations for applications such as speech recognition and image processing. Recently, digital IMC has been proposed to reduce data transfer energy, while still allowing for higher bitwidths and accuracies necessary for many workloads, especially with technology scaling [1,2]. Privacy of ML workloads can be exploited with physical side-channel attacks (SCAs) or bus probing attacks (BPAs) [3] (Fig. 1). While SCAs correlate IC power consumption or EM emissions to data or operations, BPAs directly tap traces between the IC and off-chip memory. The inputs reflect private data collected on IoT devices, such as images of faces. The weights, typically stored off-chip, reveal information about proprietary private training datasets. This work presents the first IMC macro protected against SCAs and BPAs to mitigate these risks., National Science Foundation (NSF), MIT-IBM Watson AI Lab, MathWorks Engineering Fellowship

Point cloud has been highly attracting the attention of the research community, due to their numerous applications in 3D computer vision. While learning-based approaches for point cloud problems have achieved impressive progress, generalization to unknown testing environments remains a major challenge due to the large discrepancies of data captured by different 3D sensors. Existing methods typically train a generic model and the same trained model is applied on each test instance. This could be sub-optimal since it is difficult for the same model to handle all the variations during testing. In this thesis, we propose novel frameworks for point cloud problems that adapt the model in an instance-specific manner during inference. Our model is trained using a meta-learning scheme to provide the model with the ability of fast and effective adaptation at test time. First, we consider the problem of point cloud registration. The objective is to estimate the 3D transformation that aligns a pair of partially overlapped point clouds. Next, we investigate the point cloud upsampling problem. In this setting, the goal is to generate high-resolution point clouds from sparse point clouds. Experimental results demonstrate the effectiveness of our proposed frameworks in improving the performance of state-of-the-art models and achieving superior results.

Contrastive Vision-Language Pre-training (CLIP) based point cloud classification model has added a new direction in the point cloud classification research domain. In this thesis, we propose two novel methods for CLIP-based point cloud classification. First, we propose a Pretrained Point Cloud to Image Translation Network (PPCITNet) that produces generalized colored images along with additional salient visual cues to the point cloud depth maps for CLIP based point cloud classification. In addition, we propose a novel viewpoint adapter that combines the view feature processed by each viewpoint as well as the global intertwined knowledge that exists across the multi-view features. Next, we propose a novel meta-episodic learning framework for CLIP-based point cloud classification. In addition, we introduce dynamic task sampling within the episode based on performance memory. The experimental results demonstrate the superior performance of the proposed model over existing state-of-the-art CLIP-based models on ModelNet10, ModelNet40, and ScanobjectNN datasets.

The ever-increasing dependence on the internet for the performance of human functions blurs some distinctions between physical and virtual worlds. Tasks like the performance of surgical procedures in hospitals depend on digital tools for efficient and effective health delivery. Still, society is witnessing both new forms of violence and the transposition of violent forms from the physical world to the cyber world. Cyberspace provides an easier option for harm to be caused to individuals because both state and non-state actors can extend their actions beyond their physical reach; commercial spyware is being deployed against political opponents in many countries. Without obtaining express consent, some organizations may be involved in trading the personal data of clients for business gains. During all these, the search for cyber peace has become difficult because of divergent views on what constitutes cyberviolence and the potency of cyber weapons. This research seeks to integrate discussions among scholars and the perspectives of some cybersecurity practitioners on cyber peace and violence. The decision to interrogate cyber peace and violence from the perspectives of cyber-security practitioners will contribute towards building some stability for these evolving concepts within the peace and conflict doctrines. Four cybersecurity professionals were interviewed on the subject. The basic human needs theory, human rights, and social justice theories are used to interrogate the understanding of cyber peace and violence. The results indicate that cyber harms targeting both state and non-state actors and installations should be considered in conflict analysis. This approach helps to enhance the concept of positive and negative cyber peace as possibilities.

We designed, prototyped, and deployed a novel companion robot for private in-home use that aims to promote engagement to support people living with loneliness. We identified and explored three key approaches to encourage engagement and provide comfort and prototyped a novel robot aiming to embody these principles: this resulted in SnuggleBot, a novel robot that is physically comforting, socially engaging, and requires care, to provide structure and increase engagement. We deployed our prototype unsupervised for a minimum of 7 weeks (with optional longer involvement up to 6 months) into the homes of seven people who live alone and self-identified as lonely. We reflect on our specific designs and how they promoted engagement and companionship. Our results indicate that robot designs incorporating physical comfort, social engagement, and requiring care have the potential to promote companionship, with many participants showing signs of bonding with the robot. Further, our design strategies were generally successful in that they promoted the behaviours and reactions that we intended. We found that most participants expressed that the robot is comfortable. The robot also promoted animism and engagement. Additionally, participants reported wellbeing benefits because of caring for the robot. This supports future research into robots developed with our design strategies, which can leverage our results to improve on our implementations of the design strategies.

An assertion is commonly used to validate the expected program’s behavior (e.g., if the returned value of a method equals an expected value) in software testing. Although it is a recommended practice to use only one assertion in a single test to avoid code smell (e.g., Assertion Roulette), we observe that it is common to have multiple assertions in a single test. One issue with tests that have multiple assertions is that when the test fails at an early assertion, the test will terminate at that point, and the remaining testing code will not be executed. This, in turn, can potentially reduce the code coverage and the performance of techniques that rely on code coverage information (e.g., spectrum-based fault localization). We refer to such a scenario as early test termination. Understanding the impact of early test termination on test coverage is important for software testing and debugging, particularly for the techniques that rely on coverage information obtained from the testing. In this study, we investigated 207 versions of 6 open-source projects. We found that a non-negligible portion of the failed tests (19.1%) is early terminated due to assertion failure, which leads to the skipping of 15.3% to 60.5% of the test code on average, and a negative impact on testing coverage. To mitigate early test termination, we propose two approaches, i.e., Trycatch (adding a try-catch block surrounding an assertion) and Slicing (slicing a test into a set of independent sub-tests, in which only one assertion and its dependent code are contained). After applying our approaches, the line/branch coverage get improved in 55% of the studied versions. Moreover, Slicing improves the performance of SBFL by 15.1% and 10.66% in terms of Mean First Rank (MFR) for Ochiai and Tarantula, respectively. We also provide actionable suggestions to prevent early test termination, and approaches to mitigate early test termination if it already exists in their project.