12023 research outputs found
Sort by
Understanding and Improving Reliability of Inference Dynamics in Deep Neural Networks
Get PDFReliability is a crucial aspect for the successful deployment of deep learning systems across various domains. In generative modeling, it is essential to create content that adheres to specific rules. In the field of control, ensuring that robots operate safely without falling or entering hazardous areas is paramount. Similarly, in visual perception, the robustness of perception results against perturbations are vital.
In this thesis, we explore the reliability of inference dynamics in deep neural networks such as ResNet, neural Ordinary Differential Equations (ODEs), and diffusion models.
We begin by examining the inference dynamics in standard networks with a discrete sequence of hidden layers, applying self-consistency and local Lipschitz bounds to enhance robustness against input perturbations. Our exploration then extends to neural ODEs, where the neural network specifies a vector field that continuously transforms the state. We employ forward invariance to achieve robustness, marking the first instance of training neural ODE policies with non-vacuous certified guarantees. The focus shifts next to diffusion models and their inference processes, particularly in adhering to symbolic constraints. For this, we introduce a novel sampling algorithm inspired by stochastic control principles. This algorithm not only guides these models in generating rule-specific content but also sets a new benchmark in symbolic music generation. Our work offers a cohesive understanding of inference dynamics in various deep learning architectures and propose new algorithms to significantly improve their reliability.</p
Complexity of Transcriptomic Data Analysis and Implications for Biological Discovery
Get PDFOver the past decade, the advancement of ‘omics’ technologies has ushered in a new era for the life sciences. Given the high-throughput nature of omics technologies, this era is characterized by unique computational challenges pertaining to data size and dimensionality, and technical and biological noise. Concurrently, it offers opportunities, as global, untargeted, and parallel measurement of large amounts of information often captures unexpected insights.
This thesis describes challenges inherent to the omics era of life sciences, particularly highlighting the increasing importance of merging expertise in biology and computer science. It describes the development of multiple software tools designed to address several of these challenges, which were immediately adopted and widely implemented in transcriptomics and proteomics research. Additionally, it contains three chapters focused on unraveling previously unquantifiable information, including the interpretation of sequencing data from organisms with low-quality reference genome assemblies and workflows for identifying novel viruses using single-cell RNA sequencing data already massively generated in research, healthcare, and agriculture.</p
Exploring the Photophysics and Reactivity of Nickel–Bipyridine Cross-Coupling Catalysts
Get PDFNi(II)–bipyridine (bpy) aryl halide complexes have been prized for nearly a decade for their catalytic potency to facilitate cross-coupling reactions. To achieve these transformations, the energy from light is leveraged to drive the key catalytic processes. Thus, Ni-mediated photoredox catalysis provides an attractive and sustainable means to replace precious metal catalysts. However, precise mechanistic information regarding how these transformations occur is limited. This thesis thus focuses on a dual experimental and computational analysis of Ni(II)–bpy aryl halide complexes and their photoproducts to provide insight into the specific photophysical and chemical pathways that these catalysts undertake for cross-coupling reactions. The first chapter is a review of the proposed mechanisms presented for Ni-mediated photoredox catalysis. Therein, certain portions of this work are also summarized. The second chapter provides a computational description of the Ni(II) excited states. The third chapter expands on this analysis with experiment, elucidating the photophysical pathway that grants entry into dark Ni(I)/Ni(III) catalytic cycles. Together, chapters two and three show that Ni(II)–bpy aryl halide complexes form low-valent Ni(I)–bpy halide species by an aryl-to-Ni ligand-to-metal charge transfer. Chapter four outlines a method to generate and study these reactive Ni(I)–bpy halide intermediates, identifying their mechanism of C(sp2)–Cl bond activation as nucleophilic aromatic substitution, tunable via the energies of the 3d-orbitals and the effective nuclear charge of Ni. The final chapter finds that these low-valent Ni species are competitive light-absorbers, and it presents a study into their ultrafast photophysics, marking the first of its kind on any Ni(I) complex. The excited-state relaxation dynamics of Ni(I)–bpy halide complexes are well described by vibronic Marcus theory, spanning the normal and inverted regions as a result of simple changes to the bpy substituents. Altogether, these studies have provided a framework to gain electronic structural control over Ni-meditated photoredox catalysis and, thus, guides the use of photonic energy as a sustainable alternative to precious metal catalysis.</p
Signal Processing for Large Arrays: Convolutional Beamspace, Hybrid Analog and Digital Processing, and Distributed Algorithms
Get PDFThe estimation of the directions of arrival (DOAs) of incoming waves for a passive antenna array has long been an important topic in array signal processing. Meanwhile, the estimation of the MIMO channel between a transmit antenna array and a receive antenna array is a key problem in wireless communications. In many recent works on these array processing tasks, people consider millimeter waves (mmWaves) due to their potential to offer more bandwidth than the already highly occupied lower-frequency bands. However, new challenges like strong path loss at the high frequencies of mmWaves arise. To compensate for the path loss, large arrays, or massive MIMO, are used to get large beamforming gain. It is practical due to the small sizes of mmWave antennas. When large arrays are used, it is important to develop efficient estimation algorithms with low computational and hardware complexity.
The main contribution of this thesis is to propose low-complexity DOA and channel estimation methods that are especially effective for large arrays. To achieve low complexity, three main aspects are explored: beamspace methods, hybrid analog and digital processing, and distributed algorithms. First, a new beamspace method, convolutional beamspace (CBS), is proposed for DOA estimation based on passive arrays. In CBS, the array output is spatially filtered, followed by uniform decimation (downsampling) to achieve dimensionality reduction. No DOA ambiguity occurs since the filter output is represented only by the passband sources. CBS enjoys the advantages of classical beamspace such as lower computational complexity, increased parallelism of subband processing, and improved resolution threshold for DOA estimation. Moreover, unlike classical beamspace methods, it allows root-MUSIC and ESPRIT to be performed directly for uniform linear arrays without additional preparation since the Vandermonde structure is preserved under the CBS transformation. The method produces more accurate DOA estimates than classical beamspace, and for correlated sources, better estimates than element-space.
The idea of hybrid analog and digital processing is then incorporated into CBS, leading to hybrid CBS for DOA estimation. In hybrid processing, an analog combiner is used to reduce the number of radio frequency (RF) chains and thus hardware complexity. Also for lowering hardware cost, the analog combiner is designed as a phase shifter network with unit-modulus entries. It is shown that any general (arbitrary coefficient) CBS filter can be implemented despite the unit-modulus constraints. Moreover, a new scheme of CBS is proposed based on nonuniform decimation and difference coarray method. This allows us to identify more sources than RF chains. The retained samples correspond to the sensor locations of a virtual sparse array, dilated by an integer factor, which results in larger coarray aperture and thus better estimation performance. Besides, with the use of random or deterministic filter delays that vary with snapshots, a new method is proposed to decorrelate sources for the coarray method to work.
Next, a 2-dimensional (2-D) hybrid CBS method is developed for mmWave MIMO channel estimation. Since mmWave channel estimation problems can be formulated as 2-D direction-of-departure (DOD) and DOA estimation, benefits of CBS such as low complexity are applicable here. The receiver operation is again filtering followed by decimation. A key novelty is the use of a proper counterpart of CBS at the transmitter—expansion (upsampling) followed by filtering—to reduce RF chains. The expansion and decimation can be either uniform or nonuniform. The nonuniform scheme is used with 2-D coarray method and requires fewer RF chains to achieve the same estimation performance as the uniform scheme. A method based on the introduction of filter delays is also proposed to decorrelate path gains, which is crucial to the success of coarray methods. It is shown that given fixed pilot overhead, 2-D hybrid CBS can yield more accurate channel estimates than previous methods.
Finally, distributed (decentralized) algorithms for array signal processing are studied. With the potential of reducing computation and communication complexity, distributed estimation of covariance, and distributed principal component analysis have been introduced and studied in the signal processing community in recent years. Applications in array processing have been also indicated in some detail. In this thesis, distributed algorithms are further developed for several well-known methods for DOA estimation and beamforming. New distributed algorithms are proposed for DOA estimation methods like root-MUSIC, total least squares ESPRIT, and FOCUSS. Other contributions include distributed design of the Capon beamformer from data, distributed implementation of the spatial smoothing method for coherent sources, and distributed realization of CBS. The proposed algorithms are fully distributed since average consensus (AC) is used to avoid the need for a fusion center. The algorithms are based on a finite-time version of AC which converges to the exact solution in a finite number of iterations. This enables the proposed distributed algorithms to achieve the same performance as the centralized counterparts, as demonstrated by simulations.</p
Superionic Conduction of Next-Generation Mobile Ions in Solids Enabled by Coordinating Ligands
Get PDFAdvancements in battery technologies are a critical step towards meeting the growing demand for sustainable energy storage solutions. The development of next-generation battery technologies using "beyond-Li" ions, like Na⁺, K⁺, Mg²⁺, Ca²⁺, Zn²⁺, and Al³⁺, could potentially offer improved performance, safety, and cost-effectiveness over traditional lithium-ion systems. However, the realization of next-generation battery technology based on "beyond-Li" mobile ions is limited, in part, due to a lack of understanding of solid state conduction of next-generation ions, which governs ion transport in electrodes, interphases, and solid electrolytes. “Beyond-Li” ions tend to have relatively low mobility in solids due to: (1) the larger ionic radii (Na⁺, K⁺, Ca²⁺), which limit the accessible migration pathways, or (2) higher charge densities (Mg²⁺, Zn²⁺ Al³⁺), which results in strong electrostatic interactions within the solid.
This work discusses several structure-property relationships and structural modifications that are hypothesized to lead to facile conduction of next-generation working ions. A notable discovery is the superionic conductivity of ZnPS3 after exposure to humid environments. Water is introduced into the grain boundaries, thereby enabling Zn²⁺ ions from the material to migrate and conduct freely in the network of adsorbed water. The introduction of water leads to potential H⁺, therefore a methodology for decoupling the contributions of Zn²⁺ and H⁺ in mixed ionic conducting solids using ion-selective EIS, transference number measurements, and deposition experiments is established.
Further extending this approach, superionic conductivity of other next-generation ions in electronically-insulating inorganic solids is achieved by leveraging the established ion exchange/intercalation mechanism of MPS3 (M = Cd, Mn) materials. The mobile cations that are introduced are coordinated with H2O ligands which simultaneously increase the size of the bottlenecks within the migration pathway and screen the charge-dense ions resulting in high mobilities. Potential applications can be extended to water-incompatible systems by replacing the water ligands with aprotic molecules.
These insights contribute significantly to the understanding and development of next-generation battery technologies, representing an important step toward the development of more sustainable and efficient energy storage solutions.</p
Expanding Frontiers in Biomedical Imaging and Synthetic Biology: Dynamic Acoustic Reporter Gene Imaging and Ratio-Tuning of Mammalian mRNA Polycistronic Expression
Get PDFThis thesis presents a comprehensive exploration of the next generation of mammalian Acoustic Reporter Genes (mARGs), unveiling a novel approach for non-invasive, real-time imaging of cellular processes and gene expression within live animals1. Building on the foundational work of first-generation ARGs2,3, which introduced the groundbreaking concept of using gas vesicle (GV) genes as genetically encoded ultrasound contrast agents, this research tackles the inherent limitations of these pioneering systems. The first segment details the development and characterization of the second-generation mARGs which significantly improve upon their predecessors by offering robust expression without the need for monoclonal screening, dynamic non-destructive imaging capabilities, and customizable acoustic properties through gene and protein level modifications. This advancement not only enhances the utility of mARGs in biomedical imaging but also paves the way for their application in novel therapeutic monitoring strategies, as exemplified by real-time tracking of tumor development and ultrasound-guided tumor biopsies that leverage gene expression information.
Further, the thesis delves into the structural, genetic, and biochemical principles underpinning GV assembly, addressing a critical knowledge gap that has persisted despite the utility of GVs in ultrasound imaging. Understanding these assembly mechanisms is crucial for the engineering of improved ARGs.
The exploration then extends into innovative bioengineering methodologies, specifically Stoichiometric Expression of Messenger Polycistrons by Eukaryotic Ribosomes (SEMPER), a synthetic biology breakthrough enabling the expression of multiple proteins at precise stoichiometries from single, compact transcripts.4 SEMPER represents a strategic advancement in the field, facilitating efficient formation of multi-protein complexes, minimizing cellular toxicity, and broadening the scope of potential applications in genetic engineering, including the creation of enhanced cell lines and circuits for research and therapeutic purposes.
Collectively, this work not only advances our understanding of GV-based ultrasound imaging and gene expression tracking but also introduces versatile genetic tools for the manipulation of cellular machinery. These achievements mark significant strides in the fields of synthetic biology and molecular imaging, setting the stage for future innovations in non-invasive diagnostics, cellular therapy, and cancer monitoring research. Through the integration of improved acoustic reporter genes, insights into gas vesicle assembly, and the SEMPER method for gene expression, this thesis embodies a holistic approach to overcoming current challenges and unlocking new potentials in biomedical engineering and synthetic biology.</p
Hydrogen Incorporation in Rutile- and Perovskite-Structured Minerals and Their Analogues
Get PDFFor several decades now, it has been known that large quantities of hydrogen can be stored in the earth’s mantle. This hydrogen, which is disseminated as defect components in nominally anhydrous minerals (NAMs), can have an outsized influence on minerals’ bulk properties, potentially impacting planetary-scale processes. However, a description of how this hydrogen is sequestered in NAMs — its distribution between phases, its inhomogeneity between different mantle regimes, and the variety of defects involved — has evolved significantly with time. Deciphering hydrogen’s role in the deep earth requires a detailed understanding of how hydrogen incorporates into mantle phases, beginning at an atomistic and structural level. Unfortunately, for a variety of reasons, directly measuring the crystallographic positions of hydrogen in most NAMs represents an exceptionally high technical barrier. Thus, hydrogen’s structural state is, in many phases, incompletely understood. One approach for addressing this is to incorporate the use of computational methods like density functional theory (DFT) in the interpretation of analytical methods that can provide indirect structural information, like Fourier transform infrared spectroscopy (FTIR). This is the methodology employed by the work outlined in subsequent chapters.
This thesis focuses on two specific mineral structures found within the deep earth — the rutile and perovskite structures — and explores some of the many possible hydrogen defect states in these phases. These include not only the conventionally considered hydroxyl (OH⁻) group, but also hydride (H⁻), an anionic form of hydrogen whose role in the mantle has yet to be considered in detail. The predictive and interpretive capabilities of DFT are utilized in studies on stishovite, rutile-type TiO₂, SrTiO₃, and davemaoite to both elucidate hydrogen’s incorporated state in these phases and make predictions about yet-to-be-observed hydrogen defects. Detailed spectroscopic studies on rutile-type TiO₂ and SrTiO₃ perovskite provide new insights into both hydrogen and non-hydrogen related defect structures in these materials, with implications for future studies of NAMs.</p
Topological Phenomena in Time-Multiplexed Resonator Networks
Get PDFIn 2008, the prediction that gyromagnetic photonic crystals could host analogs of the quantum Hall effect sparked a revolution in photonics, as it became apparent that the synergy between photonics and topological physics provides distinct opportunities for fundamental research and technological innovation. Since then, topological photonics has produced experimental realizations of numerous theories from topological condensed matter physics, while the inherent robustness of topological edge states has enabled novel devices like topological lasers and topological quantum sources. Despite this success, practical challenges limit the breadth of topological phenomena accessible to the existing experimental platforms for topological photonics. Therefore, to accelerate the pace of scientific discovery and to inspire the next generation of topological technologies, it is desirable to develop a platform that overcomes the limitations of traditional topological photonic architectures. In this thesis, I propose time-multiplexed resonator networks as a next-generation platform for topological photonics, and I present three experimental projects that demonstrate the diverse capabilities of this platform.
In the first project, I use a time-multiplexed resonator network to demonstrate topological dissipation, in which nontrivial topology is encoded in the dissipation spectrum of a resonator array. I show measurements of dissipative topological phenomena in one- and two-dimensions and discuss how topological dissipation can be used to design resonator arrays with topologically robust quality factors. In the second project, I adapt a time-multiplexed resonator network to realize a topological mode-locked laser, and I show that this laser can realize non-Hermitian topological phenomena that had not previously been demonstrated in topological photonics. Finally, I experimentally study the dynamics of cavity solitons in a topological resonator array. This project demonstrates a general technique for realizing cavity solitons in large arrays of coupled resonators, which has become a relevant challenge in the soliton community over the past several years.</p
Sampling the Evolution of Solar System Cometoids
Get PDFComets are commonly defined to be planetesimals visibly losing mass through volatile sublimation. In the solar system, such behavior characterizes but a brief stage in the overall evolution of these objects, as limited by their supply of accessible volatile materials and the often short dynamical lifetimes of orbits sufficiently near the Sun for said volatiles to actually volatilize. In this thesis, I explore the characteristics of several different types of ``cometoids''---planetesimals visibly exhibiting comet-like mass loss sometime in their recent past, present, or near future---in both the outer and inner solar system at different stages in their physical and dynamical evolution. I first use stellar occultations---or rather, the lack thereof detected---to constrain their abundance of kilometer-scale objects in the Kuiper Belt, from which many comets are sourced. I then evaluate how the optical brightness, color, and polarization of dust ejected by a classical, currently active comet changes when exposed to the space environment in order to probe the material properties of its nucleus. Finally, I investigate an otherwise ordinary but active asteroid to explore how intense solar heating as it passes very near the Sun can volatilize its rocky surface to produce bright sodium emission explaining its comet-like behavior
Additive Manufacturing of 3D Micro-Architected Materials for Device Applications
No full textNatural cellular biomaterials typically consist of hard and soft constituent materials that are hierarchically ordered to achieve outstanding mechanical properties, e.g., light weight, mechanical resilience, multi-functionality, etc. Architected materials are a new class of engineered materials with meticulously controlled internal structures that produce properties that differ from or exceed those of their constituent materials. Recent developments in additive manufacturing offer an extraordinary opportunity to rationally design the structure and chemical composition of architected materials to optimize properties and functionalities for a wide range of device applications. Here we first present a framework that combines an artificial intelligence tool and two-photon lithography in order to design and fabricate optimal porous structure with the desired anisotropic mechanical properties. The biomimetic and extremely tunable natural of the structures generated by the framework enables the great potential to be used as the bone scaffold design strategy which meets the requirements of complex anisotropic and heterogeneous mechanical properties of the vivo environment. The designed the architectures are meticulously verified by in situ Nanomechanics. These theory-informed experiments revealed close agreement between experimental data and artificial intelligence-predicted stiffness anisotropy, which opens a pathway for uncovering previous unattainable design space of elasticity vs. 3D architecture mapping in quantifiable and deterministic way. Besides, we explore the structural and material effects of additively manufactured microrobots which is powered by external physical fields for complex therapeutic assignments. The excellent movability and controllability permit the microrobots to be used as minimal invasive instruments for precise application in healthcare. The synergistically optimized microstructures and chemical composition enables the microrobots great potential to be applied to in vivo clinical applications