545298 research outputs found
Sort by
Artificial Musical Minds: The Helmholtz Machine and Expressive Music Generation
Get PDFThis dissertation is a statement of the author’s attempts in understanding and simulating musical scenarios, and generating music with generative AI models. To investigate music perception, cognition, and music knowledge construction processes, we adopted a hierarchical feedback neural network, the Helmholtz machine, which was designed to simulate the bottom-up and top-down cortical processing pathways of human brains. In Chapter 1, we justified the prevalence of hierarchical structures in human cognition and language/music construction, and introduced the foundations, working mechanisms, and various properties of the Helmholtz machine and variational learning. Based on the structure and learning behaviors of the Helmholtz machine, we designed a targeted algorithm to enhance the model’s performance while keeping the essential qualities of the Helmholtz machine unchanged, such as simulated local synaptic efficacy, alternative parameter updating, and joint-optimization/mutual-approximation training of recognition and generative models. In Chapter 3, drawing upon our original motivations, we systematically investigate the human cognitive functions via predictive processing, motor responses via salience sampling, and the unified process of the action-perception loop theorized by the Free Energy Principle, all simulated with our Helmholtz machine model. We sublimate the research topics in music to a higher level of general cognition, perception, enactivism, group coordination, social interaction, and cultural niche construction problems, and through addressing these grander subjects, we understand music, musicking, and ethnomusicology with foundations rooted in grand humanity. Chapter 4 is an independent chapter that approaches music and AI from a different perspective. Instead of investigating the music knowledge construction conceptually, we divert to a pragmatic path that aims to generate real music performance directly. We devised an expressive data processing method that retains most of the expressiveness in MIDI-format piano performances through data pre-processing. Rather than hierarchical models, we chose the classical sequential neural network, Long-short Term Memory (LSTM), and modified it to a multi-argument form to capture the interdependencies among multiple input fields of each MIDI note. Unlike the theoretical investigations with the Helmholtz machine, this music AI generation model generates hi-fi piano performance with expressive micro-timing, rich polyphonic texture, involved musical complexity, and creative musical structure mimicking the Kontakte’s Moment Form
Lagrangian analyses of ichthyoplankton: insights into transport processes in the Southern California Bight
Get PDFPhysical dynamics in the ocean play critical roles in the early life stages of fish. Larvae and eggs of many species behave as passive drifters in the upper water column, making dispersal to or from favorable habitats an important process in their survival to the juvenile stage. Traditional fisheries oceanography is often focused on the Eulerian perspective – comparing ichthyoplankton abundances with environmental and oceanographic indexes where the fish were sampled to understand what conditions drive survival, and ultimately, lead to recruitment to the adult population. This perspective is useful for fisheries management and predicting how changing ocean conditions may affect commercially important fish. However, the Eulerian perspective is a “snapshot” view in space and time and lacks information regarding ichthyoplankton transport pathways and the associated oceanographic conditions experienced by the individual along their drifting trajectories. Here, I leverage Lagrangian methods to investigate ichthyoplankton dispersal patterns in the Southern California Bight.In Chapter 1, I used high-resolution glider-measured velocities to assess the utility of satellite-derived geostrophic currents for understanding upper ocean transport processes. I found that satellite-derived geostrophic velocities perform well at resolving ocean circulation patterns in the upper 200 m of the California Current. In Chapter 2, I leveraged a 25-year record of satellite-derived geostrophic velocities to assess transport and retention of virtual particles representing rockfish. Spatial patterns of advection revealed high retention regions in the Southern California Bight. Interannual variability in retention was correlated with the abundance of pelagic juvenile rockfish collected months later, underscoring the importance of physical processes in the early life stages of rockfish. Lastly, in Chapter 3, I built on the previous two chapters by combining measured satellite-derived velocities with in situ collections of larval rockfish and associated otolith measurements to predict regions of spawning and dispersal. Backward-in-time particle tracking revealed that high quality rockfish were born offshore of Point Conception, while forward-in-time particle trajectories highlighted the role of the Cowcod Conservation Areas in sourcing larvae to nearby regions accessible to fishing. Overall, this dissertation demonstrates the utility of satellite-derived geostrophic velocities for understanding transport processes and emphasizes the importance of mesoscale circulation patterns on the early life stages of fish
Neural Lyapunov Methods for Learning-based Control
Get PDFLearning-based control methods have demonstrated strong capabilities in solving complex control tasks in robotics. However, their broader adoption is often limited by sample inefficiency and the lack of formal guarantees regarding the stability and safety of the learned controllers. This dissertation introduces a set of frameworks that enhance the reliability and efficiency of learning-based control by extending the use of Lyapunov methods and introducing adaptive action-selection strategies.
In this thesis, we first introduce neural Lyapunov methods that jointly optimize stabilizing controllers and Lyapunov functions, providing provable guarantees for the stability of dynamical systems. These methods significantly simplify nonlinear control design and achieve substantially larger regions of attraction compared to existing control methods. We then develop algorithms that extend neural Lyapunov methods to a variety of control settings, including model-based stabilization of nonlinear systems, model-free reinforcement learning, and imitation learning from limited expert demonstrations. In all cases, the learned controllers demonstrate enhanced stability and robustness compared to existing learning-based techniques. To further address the inefficiencies of exploration in reinforcement learning, we propose novel methods based on Extremum-Seeking Control that improve action selection by leveraging local optimization signals. These approaches enable more efficient learning by adaptively directing exploration toward the most informative regions within the state space.
Taken together, the methods developed in this thesis form a principled framework for learning-based control that is both robust and scalable across simulated and real-world robotic systems. By applying neural Lyapunov techniques to ensure formal stability guarantees, and adaptive exploration strategies to improve learning efficiency, this work closes the gap between theoretical guarantees and practical deployment. As a result, it advances the safety, efficiency, and real-world applicability of learning-based control in robotics
Interpreting Energy at Historic Sites and Museums to Inspire Climate Action
Get PDFAn introduction to the following set of theme papers in this issue of Parks Stewardship Forum, “Interpreting Energy at Historic Sites and Museums to Inspire Climate Action,” which provide examples of how to connect past energy use patterns and attitudes to new ones that are more responsive to the challenges of climate change
Atmospheric Responses to Tropical Ocean Variability: From Planetary Circulations to Asian Monsoon Predictability
No full textTropical oceans are fundamental drivers of Earth's climate, modulating atmospheric circulation across a range of spatial and temporal scales. This dissertation examines how tropical ocean variability influences both large-scale circulation features and regional climate variability, focusing on two interconnected themes: (1) the dynamics of large-scale atmospheric structures, including the Hadley circulation, seasonal superrotation, and the intertropical convergence zone (ITCZ); and (2) the variability and predictability of the Asian summer monsoon, with emphasis on the anomalous anticyclone (AAC) over the northwestern Pacific and its relationships to the El Niño–Southern Oscillation (ENSO) and internal variability.
The first part investigates how tropical ocean variability interacts with large-scale dynamical structures. Using an axisymmetric single-layer model with parameterized eddy forcing, I show that the Hadley cell's response to equatorial thermal perturbations of varying width depends critically on the strength of midlatitude eddies. The tropical thermal forcing can only expand the Hadley cell in the angular momentum-conserving regime, highlighting the critical role of eddies in regulating Hadley cell variability and its sensitivity to external forcing. A separate study demonstrates that Earth's tropical upper troposphere exhibits seasonal superrotation, maintained by stationary eddy momentum flux convergence during boreal winter and suppressed by the seasonal cycle of the Hadley circulation. These results illustrate the dynamic balance between eddy forcing and cross-equatorial angular momentum transport in shaping large-scale tropical circulation. Finally, I examine how variations in Earth's obliquity affect the ITCZ's annual-mean position and seasonal migration, providing new context for the relationship between orbital forcing, the seasonal cycle, and interhemispheric energy asymmetry.
The second part focuses on the AAC and its role in the Asian summer monsoon. I show that while El Niño events consistently favor AAC formation in the following summer, La Niña events produce weaker and less predictable responses. This asymmetry is attributed to the generally longer duration of La Niña compared to El Niño and highlights the importance of the concurrent ENSO state in summer. I also identify robust non-ENSO precursors of AAC variability, including SST anomalies in the Atlantic and tropical northwestern Pacific, and downwelling oceanic Rossby waves in the Indian Ocean, using a novel coupled model experimental framework. These results clarify the drivers of AAC variability and contribute to a better understanding of monsoon predictability
Drilled Displacement Piles: Effects of Installation on Surrounding Soils and Development of a Predictive Soil Improvement Framework
No full textThis thesis investigates the geotechnical performance of Drilled Displacement Piles (DDPs) and their interaction with surrounding soils, combining field observations, machine learning, and data-driven methodologies to advance foundation design practices. DDP installation uniquely displaces soil laterally, and promotes soil densification and enhances strength characteristics in the surrounding ground. By analyzing Cone Penetration Test (CPT) results obtained before (CPTPRE) and after (CPTPOST) pile installation from 11 construction sites, significant improvements in soil resistance were observed, particularly in sandy soils. Improvements up to 4.5 times the initial resistance were recorded, with the most pronounced changes occurring within one pile diameter and under specific spacing arrangements. To evaluate and predict these improvements, machine learning models, Artificial Neural Networks (ANN), were developed using over 25 pairs of CPTPRE and CPTPOST executed at different locations around DDPs. The ANNs successfully identified key predictors, producing high-accuracy estimations of post-installation soil properties. Further, a Hybrid multi-step sensitivity-driven Evolutionary Polynomial Regression tool was employed. This approach helped predict post-installation CPT tip resistance, sleeve friction, and normalized tip resistance in the surrounding soil. The results indicated that the change in CPT-based soil resistances was most influenced by factors such as depth, initial CPT tip resistance, pile arrangement, pile spacing, and the distance between the DDP and the CPTPOST location. The model's accuracy was subsequently evaluated, and a 75% confidence interval was established. Finally, a site-specific investigation in Northern California demonstrated how variations in installation parameters significantly influenced the degree of soil improvement, underscoring the role of drilling speed and thrust in pile performance. Comparative analysis across two different zones showed that higher penetration rates led to more significant soil improvement, due to more efficient displacement. The data analysis showed that when the penetration rate fell below a threshold value, insufficient soil displacement occurred, resulting in reduced ground improvement. Collectively, this work integrates advanced analytics, field data, and foundational engineering principles to deliver practical, performance-based design tools that improve reliability, reduce overdesign, and strengthen collaboration between academia and industry in geotechnical engineering
Bureau of Land Management Conservation Lands and BLM’s Future
Get PDFThis essay addresses the past and future of the federal Bureau of Land Management (BLM). Although the least known of the four principal US federal land management agencies, it looks after the largest amount of land, about 250 million acres. Almost all are in the West and in Alaska, as shown on the following map. A growing proportion are stewarded primarily to preserve and allow the general public to enjoy their scenery, wildlife, and historic and other cultural values (hence the “Conservation Lands” of the title). Part One provides a capsule history of how this all came about, focusing first on the events leading up to BLM’s establishment in 1946, and then on the events leading up to the present. In Part Two, I assess current trends and what they suggest about the future of these BLM National Conservation Lands
The Big Picture: Achieving Landscape-scale Conservation on Public Lands
Get PDFThis paper examines the long-term, very-American pattern of participatory conservation involving public lands managed by the Bureau of Land Management (BLM) as part of large landscapes in the West. Conservation outcomes have become common, shared goals publicly expressed and supported through designations, as well as protection and restoration efforts. With varying degrees of success, communities of caring people have been the driving force underlying conservation
The National Parks and Geography (book excerpt)
Get PDFAn excerpt from the book The Parks Belong to the People: The Geography of the National Park System