42337 research outputs found
Sort by
Sequence-to-sequence linguistic frontend modelling for text-to-speech and its improvement
Get PDFWith the development of deep learning techniques, the last decade has witnessed the
transition of Text-to-Speech (TTS) from Statistical Parametric Speech Synthesis (SPSS)
to neural TTS, and more recently to large-scale neural TTS. Loosely speaking, this
evolution is caused by the increased modelling capability of the acoustic model. As
acoustic models become increasingly powerful, the burden of contextualising the input
text sequence (i.e., deriving and predicting various types of context from the text
sequence) that was typically taken by the linguistic frontend has been largely shifted to
the acoustic model. As a result, the linguistic frontend and its output pronunciation sequence
become increasingly simplified. Despite this, some recent studies have shown
that for languages with irregular pronunciation patterns (e.g., English), pronunciation
sequences remain an effective input representation for acoustic models to ensure the
pronunciation accuracy of synthesized speech. In other words, a high-quality explicit
linguistic frontend is still a necessary part of (both conventional and large-scale) neural
TTS for these languages. However, a conventional pipeline-based frontend is difficult
to build and is vulnerable to compounding errors (a.k.a., accumulated errors).
Recently, Sequence-to-Sequence (Seq2Seq) frontends have become a new paradigm
for linguistic frontends, directly converting the text sequence to a corresponding pronunciation
sequence at the sentence level. Due to unified modelling, Seq2Seq frontends
can greatly mitigate the drawbacks of pipeline-based frontends. Following this
line of research, this thesis aims to facilitate the initial building and subsequent improvement
of Seq2Seq frontends. To overcome the lack of annotated pronunciation
training targets for initialising the Seq2Seq frontend, we opt to apply a bootstrapping
method, where a pre-existing pipeline-based frontend is utilised to generate pronunciation
sequences for large amounts of unlabelled text to form the bootstrapping training
dataset for training the Seq2Seq frontend. The bootstrapped Seq2Seq frontend is
shown to achieve impressive memorisation performance (99.9% word accuracy) and
generalisation performance (82% word accuracy).
Nevertheless, the fixed lexical coverage of bootstrapping training dataset (due to
the fixed-sized pronunciation dictionary built within the pipeline-based frontend) poses
a major limitation to the bootstrapped Seq2Seq frontend. To overcome this limitation,
we opt to resort to some easily-accessible extra training source covering certain word
types (and their pronunciation knowledge) that have not yet been covered in the original
bootstrapping training dataset. Based on this, three different methods are proposed
in this work, including (i) a Forced-Alignment (FA)-based method that leverages transcribed
speech audio as an extra training source, (ii) a Multi-Accent Bootstrapping
(MAB) method that leverages bootstrapping training data of some extra accent(s) other
than the main accent we aim to model as an extra training source, and (iii) a Multi-Task
Learning (MTL)-based method that also leverages transcribed speech audio as an extra
training source, but has a simpler implementation process compared to the FA-based
method. The proposed methods are shown to be effective in acquiring novel pronunciation
knowledge of those previously uncovered word types from their corresponding
extra training sources. Our experimental results show that, for those previously uncovered
word types, the FA-based method can improve the word accuracy by more
than 3% absolute, the MAB method can improve the word accuracy by more than 12%
absolute and the MTL-based method can improve the word accuracy by more than 3%
absolute
From inauguration to upgrade: studying charmless beauty at LHCb
Get PDFThe LHCb experiment at CERN is designed for the study of b-hadrons as a laboratory to test the Standard Model of particle physics. Studies of charmless B-decays are an avenue for such studies. The B0 → ϕϕ decay is a highly suppressed process that is allowed by the Standard Model that is not yet observed at LHCb using data collected in the period 2011 - 2016 inclusive. The rare nature of this mode makes it sensitive to beyond Standard Model physics which can manifest itself in the form of an enhanced branching fraction. This thesis presents a dedicated search for this decay mode in the full LHCb dataset, collected in the period 2011 - 2018 inclusive. No significant signal is observed and a new limit is set on the branching fraction at 1.3 × 10−8 at 90% confidence level.
The topologically identical and much more abundant decays of B0 s → ϕϕ are also studied as a probe of beyond Standard Model physics in CP violation. The B0 s → ϕϕ mode occurs via loop-level penguin processes and give access to the CP-violating phase, ϕ sss¯ s . The B0 s → ϕϕ decays in particular are sensitive to any beyond Standard Model physics that may manifest in the loop contributions. This thesis presents the expected sensitivity to ϕ sss¯ s with B0 s → ϕϕ decays that may be achieved in the Upgrade I phase of LHCb, assuming 25 fb−1 of data are collected in Run 3.
Looking beyond Upgrade I and into the future, the LHCb experiment will begin Upgrade II; research and development is ongoing with a target to begin installation in 2033. Studies of the acceptance of the LHCb detector are presented; this informs the development of the Upgrade II subdetectors by setting a common working point. The effect of reducing acceptance is presented for three modes of complimentary topology. In addition, momentum parameterisation in fast simulation is used to give preliminary sensitivities to flagship measurements like the search for B0 → ϕϕ in the Upgrade II era of the experiment
Region-based deep learning methods to enhance subtle lesion detection
Get PDFThis thesis explores deep learning (DL) techniques for analysing medical images, with a particular focus on brain CT and MRI.
We begin by proposing a comprehensive semi-automatic pipeline to tackle the challenges of preparing and standardising a dataset of routinely-collected CT brain scans from the Third International Stroke Trial (IST-3) for DL analysis. Using these scans, we develop a convolutional neural network-based method to detect acute ischemic stroke (AIS) lesions and classify the affected brain side.
To address the challenge of correctly classifying subtle lesions, we introduce the Adversarial Counterfactual Attention (ACAT) framework, which addresses the limitations of traditional CNNs in medical imaging tasks where only small parts of the image are informative. ACAT employs saliency maps to obtain soft spatial attention masks that modulate image features at different scales and increases the baseline classification accuracy of lesions in brain CT scans from 71.39% to 72.55 % and of COVID-19 related findings in lung CT scans from 67.71 % to 70.84%.
We investigate the best way to generate the saliency maps employed in our architecture and propose a way to obtain them from adversarially generated counterfactual images. They are able to isolate the area of interest in brain and lung CT scans without using any manual annotations. In the task of localising the lesion location out of 6 possible regions, they obtain a score of 65.05% on brain CT scans, improving the score of 61.29 % obtained with the best competing method.
ACAT is able to identify where an image should be modified, but not exactly how to modify it to obtain a credible counterfactual. Therefore, we present a weakly supervised method for generating healthy counterfactuals of diseased images and obtaining pixel-wise anomaly maps. This approach combines Denoising Diffusion Probabilistic Models (DDPM) and Denoising Diffusion Implicit Models (DDIM) in a novel way to perform targeted modifications to pathological areas while preserving the rest of the image. The process begins with a saliency map obtained through ACAT, which approximately covers the pathological areas. A diffusion model trained on healthy samples is then employed, using DDPM to modify lesion-affected areas within the saliency map, while DDIM ensures accurate reconstruction of normal anatomy outside these regions.
The two parts are also fused at each timestep, to guarantee the generation of a sample with a coherent appearance and a seamless transition between edited and unedited parts.
We compare our approach with alternative weakly supervised methods on the task of brain lesion segmentation, achieving the highest mean Dice and IoU scores among the models considered
Stability condition on a singular surface and its resolution
Get PDFLet X be a surface with an ADE-singularity and let X̅ be its crepant resolution. In this thesis, we show that there exists a Bridgeland stability condition αₓ on Dᵇ(X)
and a weak stability condition αₓ̅ on the derived category of the desingularisation Dᵇ(X̅), such that pushforward of αₓ̅ -semistable objects are αₓ-semistable.
We first construct Bridgeland stability conditions απ∗H,β,z on Dᵇ(X̅) associated to the contraction X̅→X, generalizing the results of Tramel and Xia in [TX22],
then we deform them to a weak stability condition αₓ̅ and show that it descends to Dᵇ(X), producing the stability condition αₓ.
Finally, we study the moduli spaces of απ∗H,β,z , of αₓ̅ - , and of αₓ-semistable objects, and we show that the moduli spaces satisfy boundedness and openness,
and hence are all Artin stacks of finite type over C
InSAR Observations of Ground Deformation in Permafrost and Agricultural Regions of an Arid River Basin
Get PDFThe Shiyang River Basin (SRB) in northwestern China is facing critical water
stresses due to extensive groundwater exploitation for irrigation activities. With
catchment hydrology strongly influenced by upstream permafrost areas undergoing
rapid changes under climate change, the reliability of future water supply to
downstream areas is uncertain. This study applies Sentinel-1 Interferometric
Synthetic Aperture Radar time series analysis to characterise spatiotemporal patterns
in ground deformation linked to cryospheric and hydrological changes associated
with upstream permafrost degradation and downstream groundwater extraction.
Results show that agricultural areas exhibit rapid long-term (2015-2024) subsidence
rates of up to -23 mm/year and large seasonal amplitudes of ground motion (up to 35
mm), with seasonal cycles characterised by summer subsidence and winter uplift,
consistent with irrigation-driven groundwater depletion and recharge. In permafrost
areas, widespread subsidence and seasonal deformation was observed with strong
spatial variability indicative of long-term permafrost thaw, active layer freeze-thaw
cycles with a dependency of underlying cryospheric mechanisms on topography and
local soil conditions. The findings highlight the nature of catchment-wide
hydrological changes from anthropogenic groundwater extraction and climate-driven
permafrost degradation in the SRB, demonstrating the effectiveness of InSAR as a
monitoring tool for supporting groundwater management and tracking changes in
alpine hydrology
Roads, Riches, and Residuals: A Multilevel Analysis of Subnational Determinants of Violent Conflict in Syria (2011-2019)
No full textThis paper investigates the subnational determinants of violent conflict in Syria from 2011 to 2019. District-level, annual counts of total estimated fatalities, total conflict events, state-based violence, non-state violence, and one-sided violence are combined with two structural indices, (1) infrastructure (roads, population density, urban/built up land cover) and (2) land cover (croplands, open shrublands, barren/sparsely vegetated, savannas, cropland natural vegetation mosaics and grasslands), each derived via Principal Component Analysis (PCA). Three governorate-level human development indices (income, education, health) are residualised on the other two to isolate unique effects. Mixed effects negative binomial models with governorate random intercepts, and a constant-only zero-inflation term for non-state violence, estimate how infrastructure, land cover, and human development imbalances predict conflict intensity, controlling for year fixed effects. Spatial clustering is assessed via Moran’s I and Local Indicators of Spatial Association (LISA), confirming clustering of violence and identifying district-level hot spots. Results indicate that higher road density, urban cover, and population concentration are associated with significantly increased state-based and one-sided violence, while higher-than-expected education attainment may influence non-state violence against civilians. By modelling district-level conflict with governorate-level heterogeneity, this analysis demonstrates how disaggregated, multilevel models combined with spatial autocorrelation diagnostics can uncover the structural drivers of violence and highlight the trade-off between within- and between-unit inference when human development data is available only at coarser scales
Towards a new ecological data framework: defining pheno-data and pheno-fication through organism-centric design research
No full textThis thesis calls for reframing and reimagining data as an ecological, ethical, and political force. Responding to the limitations of dominant data paradigms, which are rooted in capitalism and anthropocentrism, it proposes an ecological framework grounded in knowing ecological data, being in more-than-human entanglements, and the doing of design practices. Through a multidisciplinary and practice-based design approach, I develop the concepts of Pheno-data and Pheno-fication: an ecological framework for conceptualising and engaging with data that centres embodied, relational, and situated meanings within plural lifeworlds.
Drawing from critical data studies, traditional ecological knowledge, and posthumanist perspectives on ecological livingness, the thesis contrasts dominant notions of data with an alternative notion I call livingness-derived data, which emerges from organisms' bodies, relationships, and ecologies. Through research on tomatoes and the fabulation workshop method, I explore how these different notions of data are politically and ethically entangled in diverse stories and worlds. The research culminates in an ecological data framework that involves Pheno-data (ontology) and Pheno-fication (epistemology) and is based on the principles of attentiveness, responsiveness, and care-full imagination, with implications for Human-computer interaction & design, data technology, and sustainability policy. Ultimately, this thesis offers a speculative yet grounded notion of data as a vehicle for more-than-human flourishing and a good planetary life
Data-efficient learning of long-horizon dexterous robot manipulation
Get PDFAcquiring complex, long-horizon manipulation skills remains a central challenge
in robotics. Mainstream learning approaches, such as imitation and reinforcement
learning, often struggle with efficiency, demanding extensive data collection,
laborious reward engineering, or failing to overcome exploration challenges in long-horizon
scenarios. Conventional data-driven methods—imitation learning and
reinforcement learning in particular—often struggle with efficiency, demanding
extensive data collection, laborious reward engineering, or failing to overcome
exploration challenges in long-horizon scenarios. This dissertation tackles these
limitations by (i) exploiting rich, high-dimensional tactile feedback, (ii) breaking
complex skills into subskills, and (iii) introducing new frameworks that leverage
privileged actions with curriculum learning, enabling the policy to efficiently
acquire long-horizon skills.
Our work first addresses data efficiency in imitation learning by leveraging
high-dimensional tactile data. We introduce a novel framework that fuses encoded
tactile information with robot’s proprioception, enabling a dual-arm system to
acquire fine-grained, contact-aware grasping skills from a very limited number
of human demonstrations. The resulting policy demonstrates emergent, robust
behaviors, such as reactive re-grasping, that were not explicitly taught.
We extend our research to efficiently learn long-horizon tasks by proposing
a haptic-based skill segmentation framework, DexSkills. By learning a library
of primitive skills from human demonstrations, our system can analyze a new,
complex task and automatically decompose it into a sequence of these known
primitives. This approach enables one-shot learning and autonomous execution
of long-horizon tasks, relying solely on proprioceptive and tactile data.
In scenarios where demonstrations are difficult to provide, we address the
challenge of inefficient exploration in reinforcement learning by introducing a
structured learning curriculum built on the concept of privileged actions. By
temporarily granting the robot physically infeasible abilities in simulation—such
as relaxed collision constraints and virtual forces—our framework drastically simplifies
the exploration problem. A curriculum gradually removes these privileges,
guiding the policy to discover robust, physically-valid solutions for complex, multistage
tasks without task-specific reward engineering. This method not only learns
challenging non-prehensile manipulation skills but also demonstrates the ability to
autonomously adapt its strategy in response to significant environmental changes
Population Mixing, Space, and Pandemic Spread: A Geo-spatial Machine Learning Analysis of Six COVID-19 Waves in the UK
Get PDFPopulation mixing behaviours (including migration and commuting) are widely regarded as an important mechanism driving the spread of viruses across regions, however, existing studies have mostly ignored their non-linear relationships and spatial heterogeneity characteristics. In this paper, a geographically weighted random forest (GWRF) model was constructed to systematically investigate the impact of population mixing on infection rates from a spatial machine learning perspective, using the six major epidemic waves of COVID-19 in the UK as the study period. The study selected mesoscale geographic units (MSOA) as the object of analysis, and controlled for socio-demographic factors such as health status, employment rate, ethnic composition, etc., and uncovered the importance distribution and non-linear marginal effects of the variables using the SHAP method. Meanwhile, the SaTScan spatial scanning statistics were used to identify the clustered areas of high prevalence of infection at each stage, and the fitting performance and residual spatial structure of the GWRF model were evaluated.
The results showed that the demographic mixing variables exhibited significant spatial heterogeneity and time-varying characteristics in different epidemic waves. Among them, migration diversity and commuting distance had strong explanatory power for the infection rate in most periods, and SHAP analysis revealed that there were threshold effects and asymmetric patterns in the influence of key variables on infection risk. This paper reveals the complex mechanism of population mixing in epidemic transmission from a spatial epidemiological perspectiv
Geospatial Analysis of the Association between Tobacco Outlet Density and Respiratory Hospitalisations in Glasgow and Edinburgh, Scotland
Get PDFThis study examines the spatial relationship between tobacco outlet density and hospitalisation rates for asthma and chronic obstructive pulmonary disease (COPD) across Glasgow City and City of Edinburgh, Scotland. The research investigates whether this relationship varies locally within the cities itself and across the cities as a whole. Using a spatial analysis approach, the study focuses on Intermediate Zones (IZs) in both cities. Significant spatial autocorrelation was observed for both asthma and COPD hospitalisation rates in Glasgow, while Edinburgh showed no significant spatial clustering. No significant association between tobacco outlet density and hospitalisation rates was found in either city. However, local-scale analysis in Glasgow revealed a negative correlation between higher tobacco outlet density and lower COPD hospitalisation rates in certain IZs in the eastern part of the city. The inclusion of socioeconomic covariates, particularly income, improved the explanatory power of the models. No significant interaction effects were observed between tobacco outlet density and environmental factors, such as PM₂.₅, NO₂, and NDVI. These findings suggest that while tobacco outlet density may influence local health outcomes, socioeconomic factors are more significant drivers of respiratory health inequalities. Further research with alternative health outcomes, such as outpatient visits, and expanding the geographical scope to areas with more environmental diversity could provide a more comprehensive understanding of the relationship
Access Repository Dashboard
Do you manage Edinburgh Research Archive? Access insider analytics, issue reports and manage access to outputs from your repository in the CORE Repository Dashboard!