143174 research outputs found
Sort by
Diagrammatic algebra for equivariant neural network architectures
No full textGroup equivariant neural networks is an area of deep learning that looks at how symmetries can be encoded in neural network architectures as an inductive bias. Historically, many group equivariant neural networks have been designed using either group convolutions or tensor power representations that have been decomposed into irreducible representations. Whilst these approaches have value, they come with drawbacks. In the former, a change of basis into the Fourier domain is often required, which can be computationally expensive. In the latter, the irreducible decomposition of a tensor power representation for most groups is unknown.
We instead establish weight tying as a fundamental paradigm for constructing group equivariant neural networks. We show that, for many important groups, including the symmetric, alternating, orthogonal, special orthogonal, and symplectic groups, we can determine the weight matrices that appear in group equivariant neural networks having tensor power representations of Rn as their layer spaces without needing to decompose them into irreducible representations. We study the combinatorics of set partition diagrams that are associated with each group to achieve a full characterisation of these weight matrices. We create a deeper structure for understanding the group equivariant neural networks themselves by developing a monoidal category theoretic framework and use it to construct a fast multiplication algorithm for the linear layer functions for four of these groups. We extend this framework to characterise the equivariant weight matrices for the automorphism group of a graph. Finally, we show that the equivariant neural network paradigm can be broadened to quantum groups, which often describe symmetries in non-commutative geometries. In particular, we use Woronowicz’s version of Tannaka–Krein duality to derive Compact Matrix Quantum Group Equivariant Neural Networks. We hope that our work will inspire others to develop neural network architectures that encode different algebraic structures as an inductive bias.Open Acces
Advances in stochastic processes with machine learning applications
No full textStochastic processes play a fundamental role in our attempt to model the world around us. This thesis explores several topics within this field, with a particular focus on prediction, estimation, and inference. We begin by addressing the empirical question of whether order book-based financial markets exhibit predictability, using deep learning architectures to extract information from their microstructural dynamics. Moving from discrete-time to continuous-time modeling, we then consider continuous-time autoregressive processes, developing consistent and asymptotically normal estimators for their drift parameters under both continuous and discrete observations. Finally, we present new limiting results for the empirical expected signature, a powerful non-parametric statistic of path-valued random variables, and review its application in various machine learning algorithms, demonstrating the practical benefits of our findings. A recurring theme across all three projects is the use of high-frequency data, an essential bridge between theory and practice.Open Acces
Fires across conservation frontiers: alternative management approaches for addressing wildfires in East and Southern African savanna-protected areas
No full textColonially inherited fire suppression policies and the expansion of exclusionary protected areas across East and Southern African savannas have engendered a wildfire paradox. Inside protected areas, efforts to eliminate fires from the landscape have paradoxically resulted in the accumulation of flammable biomass and higher intensity wildfires. Outside protected areas, rangelands are continuously reconfigured and reterritorialized, emerging as contentious biopolitical frontiers where livestock have replaced fire as the dominant fuel consumer. ‘Bottom-up’ fire management is increasingly recognised as an alternative fire management strategy to effectively and equitably address the wildfire paradox. Yet, there has been little investigation into the framing, nature, and extent of bottom-up approaches and the assumptions informing their implementation, exposing local governance structures to external appropriation and accelerating savanna degradation. This thesis interrogates current understandings of bottom-up fire management approaches across East and Southern African conservation landscapes to identify opportunities and challenges for their implementation, specifically their ability to mitigate wildfires, improve savanna health, and contribute towards resilient local livelihoods. To do this, a mixed-methodological bricolage approach was taken, consisting of two independent systematic reviews of the literature, fieldwork conducted in Kenya, and the development of a novel Belief Network model. The results show that wildfires are symptomatic of deep-rooted land and resource-based conflicts which have and continue to emerge from policy interventions supposedly designed to address inequalities. In this regard, bottom-up fire management approaches are externally constructed, governed, and reinforce ‘divide-and-rule’ policies pursued by colonial governments.Open Acces
Circular concrete through acid leaching and carbonation of cementitious materials
No full textCementitious waste, especially the fine fraction (< 4.75 mm) of waste concrete, is either downcycled or landfilled. In this research, waste concrete fines and cement paste are treated by acetic acid leaching and carbonation for upcycling. The optimal leaching conditions are determined, which maximise the Ca2+ extraction and minimise acid use and dissolution of SiO2. The recycled sand extracted under the optimal leaching conditions are comparable to virgin sand. Total replacement of virgin sand with recycled sand produces cement mortars with improved compressive strength. The silica-rich residue (SR) from acid leaching has similar or higher pozzolanic reactivity than coal fly ash. After heat treatment under 900oC, the specific surface area of SR from cement paste is reduced by 99 %. The mortars with 20% cement replacement of this material have good workability and comparable compressive strength to mortars containing traditional supplementary cementitious materials. The Ca2+rich solution is carbonated, and 99.1% pure vaterite CaCO3 is produced. Vaterite can be transformed to aragonite in a solution containing Mg2+ and Sr2+ under 60 ℃, forming a strong fibrous interlocked structure. The properties of aragonite binder can be controlled by the liquid/solid (L/S) ratio. Lower L/S ratios generate higher compressive strength. Higher L/S ratios lead to lower density, higher porosity and lower thermal conductivity. Aragonite binder can form novel lightweight load-bearing thermal insulating materials.
Currently, this process is costly and carbon positive. However, this can be improved by acid/alkali recovery, using ammonium salts and developing products with higher value. Based on laboratory results, this process can achieve a net CO2 sequestration of 93.8 kg/tonne of waste concrete with the use of sustainable acetic acid and ammonia and zero-carbon energy. This research demonstrates the potential for the scaled-up application of indirect carbonation and the development of circular concrete.Open Acces
The role of the neighbourhood food environment in obesity and dietary intake. Evidence from the South Asia Biobank
No full textBackground: More than 77% of all non-communicable diseases (NCDs) deaths occur in lowand middle-income countries (LMICs). Obesity and unhealthy diet are leading risk factors associated with NCDs. NCDs are caused by exposure to factors in our environment that are amendable to policy interventions, with interventions targeting food environments highlighted as potentially effective for population-wide improvements in diets and weight status. This thesis explores the influence of the availability and advertising mechanisms in the neighbourhood food environment on diet, obesity and their associated inequalities among different population groups (by income, sex, and urbanisation) in the context of LMICs. Methods: I used two datasets (surveillance and environmental mapping) from the South Asia Biobank, a cross-sectional study of adults in Bangladesh, India, Pakistan, and Sri Lanka. I geospatially linked individual-level data with neighbourhood food environment characteristics (number of food retailers and food products advertised) around participants' residential addresses. Global Human Settlement Layer data was utilised to define urbanisation. Using multivariate regression models, I studied associations between the neighbourhood food environment and obesity and diet. Decomposition models were used to evaluate whether differences in the neighbourhood food environment contribute to the urban-rural inequalities in obesity and diet. Findings: The availability of unhealthy food retailers and advertising of unhealthy food in the neighbourhood were significantly associated with unhealthy dietary patterns and higher likelihood of obesity, especially among females and low-income groups. The availability of healthy food retailers was associated with healthier diets and lower likelihood of obesity, especially among females and high-income groups...Open Acces
Linear and non-linear methods of source separation in neurophysiological time series
No full textBlind source separation of neurophysiological time series into its constituent neural activity is a key component of modern concepts of neural interfacing and an important tool for motor neuroscientists. With concurrent rapid advances in high-density electrode arrays, there is now a real possibility of devices driven directly by the nervous system leaving the lab and becoming a core part of machine interaction in everyday life. Such an advance would represent a leap forward in a diverse range of technologies, including intuitive prosthetics, neurological diagnostics, and consumer device. However, contemporary linear algorithms can return poor results when subjected to some of the signal contaminants that will be faced in the wild, such as high or varying noise levels and the non-stationary effects of dynamic contraction. New approaches are needed if source separation routines are to be robust enough for practical use.
One clear avenue of enquiry can be found in the recent explosion in techniques for training non-linear functions such as deep neural networks. Deep neural networks are highly flexible and have proven efficacy on a variety of data types, including many time series applications. However, their adoption would significantly increase the complexity of the blind source separation pipeline, so their presence would need to be well justified. In this thesis I explore a range of deep learning methodologies, and in particular how they can be effectively blended with contemporary linear algorithms. I demonstrate that such hybrid approaches can bring a range of advantages, such as improved management of noise, the automatic identification of incorrect timestamps and the ability to compensate for non-stationary effects in the signal.Open Acces
Increasing extreme rainfall and rapid urbanisation major drivers behind Gaborone’s deadly floods
No full textPlanar shock-induced bubble collapse and jetting in water captured via x-ray phase contrast imaging
No full textShock wave-bubble interactions in water manifest rich dynamics driven by a combination of strong pressure and density mismatches. They have a wide variety of applications, including the injection of pharmaceuticals, and through scaling, enable the exploration of various aspects of high-energy-density systems such as inertial confinement fusion. In this work, the interaction between a micrometric nitrogen bubble and a planar shock wave, characterized by a Mach number of M = 1.24 and a peak pressure of p max = 0.57 GPa, is experimentally recorded using ultra-high-speed x-ray phase contrast imaging. Highly resolved radiographs provide access to all phase discontinuities along the beam path, offering quantities such as the time-varying bubble size, the speed of a jet produced during the bubble collapse, and the time evolution of the shock wave front, which are critical benchmark data for numerical scheme validation. This study addresses the lack of well-characterized, repeatable, and high spatiotemporal resolution experiments at negative Atwood numbers by providing shock-bubble visualization and corresponding numerical simulation
Vehicle-to-vehicle connectivity for real-time traffic incident detection on motorways: a traffic microsimulation study
No full textThe rapid pace of urbanisation and advancements in Intelligent Mobility (IM) bring both opportunities and challenges. While improved mobility options enhance convenience, they also contribute to traffic congestion and increased incident risks. Innovations in artificial intelligence—such as machine learning, big data, and image recognition—have transformed vehicles from mechanical systems into intelligent, connected entities. Connected Vehicle (CV) technology, which includes vehicle-to-vehicle (V2V) and vehicle-to-infrastructure (V2I) communication, enables real-time data exchange, improving traffic safety and efficiency.
Addressing traffic congestion and incidents—two key transport issues—requires advanced solutions. Incident Detection (ID) algorithms, supported by V2V connectivity, offer real-time alerts to road users, helping mitigate congestion and reduce accident risks. This study developed a V2V-based ID algorithm, evaluated using indicators like vehicle delay, queue length, macroscopic fundamental diagrams (MFDs), and conflict numbers. Using VISSIM traffic microsimulation and real-world data from the M1 motorway, scenarios with varying vehicle types and CV market penetration rates (MPRs) from 0% to 100% were assessed.
Simulations revealed that CV integration significantly improves traffic flow and safety. At 100% MPR, average delays dropped by 50%, and conflicts decreased by 55–70%, depending on incident duration. Five incident scenarios with different lane closures and durations further demonstrated that higher CV penetration leads to reduced disruption and better traffic performance.
The study also analysed 2019 traffic data on the M1 to explore incident and congestion patterns, revealing notable correlations. The findings support the potential of CV-enabled incident detection to enhance real-time traffic management and safety. This research provides a robust framework for integrating CV technology into existing infrastructure, aiding transport planners and policymakers in reducing incidents and improving congestion strategies.Open Acces