Heriot-Watt University
ROS: The Research Output Service. Heriot-Watt University EdinburghNot a member yet
4689 research outputs found
Sort by
The influence of geometric properties of data distributions on artificial neural networks
Full text linkDue to recent advances in algorithmic methods, software and hardware, artificial
neural networks have reached or even surpassed human-level performance in visual
object recognition. Today, they are ubiquitous in real-life applications, such as
autonomous vehicles, health care and various industrial settings. This remarkable
achievement, however, is clouded by a significant deficit in robustness to distribution shifts, yielding concerns about their readiness to be deployed in safety-critical
applications. At their core, artificial neural networks can be viewed as progressively
disentangling complex distributions of images to make them suitable for linear separation. This perspective offers a methodology using geometric properties and methods to study their behaviour. In this thesis we use this methodology and study
several open questions that lie in the intersection between neural networks’ lack of
robustness to distribution shifts and the geometric properties of data distributions
and representations. First, we study the effect of the three main geometric properties, namely the intrinsic dimension, extrinsic dimension and entanglement, on the
sample complexity. Complementary to previous works we show a strong interdependency between intrinsic dimension and entanglement, where the intrinsic dimension
only affects the sample complexity if the entanglement of the distribution is high.
Further, we show that the entanglement of label-specific distributions is the leading
contributor to the sample complexity in general. In the second part, we investigate
the geometric complexity of decision boundaries. We show that state-of-the-art robust neural networks learn geometrically more complex decision boundaries than
standard ones which confirms a previously made hypothesis and, when combined
with the results of the first part, at least partially explains the increased sample
complexity of robust training. We also propose an upper bound on the perturbation
magnitude over which provably a geometrically more complex decision boundary is
required. Further, we show for real-world image benchmarks that our bound also restricts the introduction of label noise. In addition, we show that the commonly used
nearest neighbour distance overestimates the robust radius of complex image distributions a distributions. In the final part, we compare several different state-of-the-art robust
training paradigms and show that dimensionality reduction of their hidden representations is a common mechanism shared amongst them, despite fundamentally
different approaches to robust training. We demonstrate that part of this dimensionality reduction is due to sharing of features between semantically similar classes.
In summary, we show in this thesis that studying neural networks through the lens
of geometric properties yields practical insights into their sample complexity and
generalisation behaviour as well as the mechanisms that result in representations
that are robust to distributions shifts
Production and fluidisation of thermochemical energy storage particles
No full textTo view abstract please refer to PDF
An adaptive robot for sports and rehabilitation coaching
No full textThis thesis investigates how, and to what extent, an adaptive robotic coach could
be used to increase motivation to, and effectiveness of, long-term repetitive solo
practice in squash and rehabilitation after stroke.
Stroke is one of the leading causes of acquired adult disability with survivors commonly suffering permanent impairments such as fatigue and weakness in the arms
and legs. Although past research strongly suggests that unsupervised rehabilitation
practice is beneficial to the patient, it is often not adhered to due to (among other
reasons) a lack of motivation. Squash, on the other hand, is an intermittent, high-intensity racket sport in which repetitive, solo drills are used frequently by many
of the top professionals. However, they are used much less frequently by players of
lower levels, indicating a lack of motivation when a coach is not present. These two
use cases, from different domains, are considered in the same body of work due to
the similarities in the individual, often unsupervised and repetitive nature of practice that helps in making long-term functional improvements after stroke and helps
high performance sports players improve their skill level.
This thesis contributes a cross-domain implementation of a robotic coach using
the Pepper robot, which incorporates high-level personalisation to groups of users
and low-level adaption to individuals over time. The development process combined quantitative data from systematic observations and qualitative recommendations from domain professionals during semi-structured interviews, with mathematical modelling and computation techniques to produce coaching policies usable
for robotic control. Short-term evaluations in squash and stroke rehabilitation validated the novel cross-domain design and implementation process and revealed that
the robotic coach was viewed by non-professional squash players as more interesting/enjoyable, more socially competent, and a more effective coach than a robot
that didn’t offer any coaching behaviours. Finally, a long-term evaluation showed
that Pepper was able to make solo squash sessions more interesting/enjoyable, gave
users a higher sense of perceived choice, and allowed players to make larger technical
improvements within sessions than during regular, unsupervised practice.Engineering and Physical Sciences Research Council (EPSRC), Grant ID: EPSRC DTP1
Indenyl/fluorenyl-N-heterocyclic carbene rhodium complexes for catalytic C-H activation and borylation
Full text linkThis work outlines the synthesis of several new η
5
-indenyl rhodium(I) N-heterocyclic
carbene (NHC) cyclooctene (COE) complexes. The effect of changing the NHC from
SIDipp to IDipp, IMes, SIMes or IMe4 was investigated. Changing the ligand was found
to influence both the steric profile and electronic properties of the complex. The
electronic properties of the system were investigated by the facile synthesis of the
carbonyl analogues by addition of carbon monoxide to the cyclooctene complexes. This
allowed for an indirect method of understanding the electronics of the complex through
IR spectroscopic investigation of the carbonyl stretching frequencies. The monodentate
complexes were found to have an absorption band which extends into the visible region
of the UV-vis spectra. From experiments with blue LED light, it was found that the
cyclooctene ligand was lost under irradiation. These complexes were able to undergo
cyclometallation to form a range of hydridic products, as indicated by signals in the 1H
NMR spectra, apart from the IMe4 complex which was unable to undergo this process
due to the strained configuration of any formed product. The Dipp-substituted complexes
were able to form cyclooctane (COA), tentatively assigned to the irreversible formation
of an η
2
-alkene complex.
Several tethered complexes were also synthesised including the unsaturated methyl-substituted fluorenyl-tethered NHC rhodium cyclooctene complex and the saturated
mesityl-substituted fluorenyl-tethered NHC rhodium cyclooctene complex. The indenyl
analogue of the latter was also synthesised, however, these complexes were very low
yielding due to reduction of rhodium in these reactions. Work towards the fluorenyl and
tetrahydrofluorenyl rhodium analogues of the monodentate indenyl complexes
demonstrated that these complexes are much less stable compared to their indenyl
analogues.
The monodentate and tethered complexes were both able to catalytically borylate arenes
and alkanes with B2pin2. Arene borylation occurred at 80°C for benzene whereas alkane
borylation required higher temperatures of 140°C. Reaction profiles were constructed for
the borylation of benzene catalysed by the monodentate complexes that showed a
pronounced ligand effect, both in the length of time for the induction period required to
generate the active species and in the rate of reaction when catalysis commenced. The
indenyl rhodium cyclooctene complex bearing a SIDipp NHC ligand was confirmed to
be the fastest catalyst, predominantly due to a very short induction period, and even outperformed [RhCp*(C2H4)2] for the catalytic borylation of benzene with B2pin2.
However, the more challenging catalytic C-H borylation of alkanes with these complexes
was unfortunately very low yielding with the reaction stalling at ca. 20% yield of
decylBpin, attributed to catalyst decomposition at the temperatures required for this
reaction. Work towards first-row transition metal analogues of [Rh(Ind)(NHC)(COE)] is
also briefly discussed
Full-field reservoir simulation and optimisation with advanced well completions
Full text linkThe aim of this thesis is to develop an optimisation workflow that could be used by
reservoir engineers to evaluate the full range of advanced well completions (AWCs),
(e.g., passive inflow control device (PICD), Autonomous inflow control device (AICD)
and interval control valve (ICV) completions) and identify their optimum design in large
full-field models with limited computational resources.
An AWC optimisation workflow, which involves the development of novel proactive and
reactive optimisation methods, is introduced. The developed proactive optimisation
method couples fit-for-purpose upscaled reservoir models with efficient global
optimisation algorithm to improve optimisation efficiency and accuracy when performing
AWC optimisation for large, real-field cases. The developed reactive optimisation
method uses an economic-based approach to provide real-time reservoir management for
ICV completion in production wells.
The proposed workflow has been tested on several case studies developed with AWCs.
These cases have been investigated with different completion strategies (e.g., PICD,
AICD and ICV completions), and ICV control frequencies.
In all investigated cases, the results show that the proposed workflow yields higher NPV
gains compared to conventional completions (No AWCs). More importantly, the
proposed workflow has significantly improved the optimisation efficiency when solving
computationally intensive proactive ICD and ICV optimisations
Assessing subchondral bone strength : computational and experimental insights from micro- to macroscale
Full text linkOsteoarthritis (OA) is a multifaceted joint disease which poses significant socioeconomic burdens. Managing the burden of OA demands continual improvements to treatment strategies. Computational modelling, for example finite element modelling, could be a complementary asset here. The predictive accuracy
of such models, however, relies on an accurate mechanical description of tissues
that comprise the joint. In this thesis, we explore the micromechanical impact of
OA on subchondral bone.
Microscale elasticity and strength of subchondral bone was assessed by compression of micropillars fabricated using ultrashort pulsed laser ablation. A numerical model incorporating the laser-tissue interaction, Raman spectroscopy,
and energy dispersive X-ray microanalysis all indicated the bone tissue that comprised the micropillars was unaffected after laser ablation.
Micropillars located in none-diseased (ND), and OA subchondral bone were
compressed quasi-statically. In silico micropillar compression equipped with an
elastoplastic material model was used to back-calculate the elastic modulus and
strength of the constituent tissue. Elastic modulus remained unchanged between
ND and OA subchondral bone, whereas strength increased in OA subchondral
trabecular bone. Pillar matched Raman spectroscopy and quantitative backscattered electron imaging revealed mineralisation is the underlying determinant of
elastic modulus and strength at the microscale.
The effect of microdamage on subchondral bone strength was investigated.
The equine athlete served as a model, where repetitive exercise induces subchondral bone microdamage accumulation. A multimodal imaging method was developed to identify a surrogate for existing microdamage in vivo. This surrogate variable was incorporated into a nonlinear constitutive model for bone tissue. Whole
bone modelling showed a significant reduction in stiffness and strength at both
material and whole bone levels due to pre-existing microdamage.
The methods and results presented in this thesis give valuable insights into
mechanisms that affect subchondral bone strength at the micro- and macroscale,
which could be of value in the development of interventions used to alleviate the
socioeconomic burdens associated with this debilitating joint disease
Integrating mindfulness and character strengths to develop 21st-century-ready leaders – a case study-mixed methods approach
Full text linkHuman Resource Management (HRM) practitioners have traditionally focused on
addressing deficiencies in leaders. Despite extensive research on 21st
-century leadership
competencies, organisations continue to struggle with outdated competency lists and a
lack of consensus. While the literature specifies 'what' qualities leaders should embody,
there is a gap in 'how' to cultivate such leaders. This study promotes a strengths-based
approach to leader development inspired by positive psychology. It explores
Mindfulness-Based Strengths Practice (MBSP), an eight-week group intervention
combining mindfulness and character strengths practices, in developing 21st
-century-ready leaders. Grounded in the Broaden-and-Build (B&B) theory of positive emotions, it
posits that MBSP participants may experience immediate benefits from positive emotions
and gradual benefits from enhanced personal resources and well-being. Using the Values
in Action (VIA) classification, the study introduces a framework using character strengths
to identify and nurture leadership competencies. Adopting a critical realism perspective,
the longitudinal Case Study-Mixed Methods (CS-MM) design involved an MBSP
intervention with working students at a German higher education institution, including an
experimental and control group. The qualitative component used thematic analysis of
participants' reflective journals to explore mechanisms enabling the development of
leadership competencies during the MBSP. Mindfulness and strengths practices increased
positive emotions like gratitude, joy, hope, and love, broadening thought-action
repertoires and enhancing resources such as optimism, emotional intelligence, resilience,
self-awareness, critical thinking, goal orientation, problem-solving, decision-making,
strategic thinking, creativity, passion for learning, curiosity, bravery, flexibility,
relationship management, communication, teamwork, empathy, and emotional
regulation. The embedded quantitative strand complemented these insights by measuring
changes in character strengths and mindfulness using non-parametric statistical tests. It
quantified the extent of these changes from baseline to the end of the eight-week MBSP
intervention compared to a control group, as well as six months post-intervention. The
alignment of qualitative and quantitative findings affirms MBSP's positive impact on
developing 21st
-century leadership competencies. This research links MBSP, positive
emotions, and leader development, offering a strengths-based framework for nurturing
21st
-century-ready leaders and valuable strategies for HRM practitioners to implement
MBSP in organisations
Mathematical landscapes : three examples of modelling ecosystems with differential equations
Full text linkThis thesis introduces three novel mathematical models for ecology, each
using deterministic differential equations to capture interactions within ecosystems. These models vary significantly in their focus and approach. The
first explores the resilience of dry rangelands under different management
strategies, and demonstrates that a strong coupling between herbivores and
vegetation enhances rangeland resilience to sudden vegetation loss without
compromising long-term productivity. The second model offers a framework
for studying the eco-evolutionary dynamics of a generic state-structured ecological community. We show that even a simple implementation of our eco-evolutionary framework accurately captures evolutionary stable strategies in
rainforest community assembly. Finally, the third model investigates the
mechanisms behind biogeomorphological patterns in the Scottish Highlands.
Our model reveals that under the influence of strong winds, heather can act
as an ecosystem engineer and self-organise into regular patterns.
Despite their differences, these models share a common goal: to provide
insights into complex ecosystem dynamics. By defining “mathematical landscapes” as representations of real-life ecosystems, the Introduction emphasises
the dual nature of models: they are inherently imperfect, yet potentially very
useful. Through a critical examination of their limitations and strengths, I argue that mathematical models, when carefully constructed and analysed, can
significantly enhance our understanding of complex ecological systems. The
Conclusion highlights the practical implications of the three models presented
in the thesis, and explores pathways for a stronger integration of mathematical
landscapes with real-world ecosystem management practices.Engineering and Physical Sciences Research Council (EPSRC) grant EP/S023291/
Towards robust open-world few-shot recognition
Full text linkIn recent years, substantial progress has been made in Few-Shot Learning (FSL),
where models are trained on small, task-specific labelled datasets. These efforts have
yielded many promising FSL approaches, significantly increasing their use in bench-marking meta-learning algorithms and sample-efficient deep neural networks. However, current FSL benchmarks often overlook practical challenges in real-world applications, limiting generalisation beyond controlled research contexts. The themes
in this thesis are relevant to critical stages of the FSL pipeline, from training and
evaluation to continual learning and navigating towards robust open-world environments. The thesis begins with an exploration of training methodologies to handle
class imbalances. Moving to evaluation, the thesis explores the complexities of open-set recognition, testing the adaptability of methods in detecting outlier classes encountered beyond the training environment. Finally, the thesis examines continual
learning scenarios, emphasising the necessity of efficient re-training and updating
pipelines to maintain and enhance performance over time as new data becomes
available. Together, this thesis navigates towards robust open-world environments,
ensuring comprehensive coverage from initial training and evaluation to continual
re-training of methods. The individual summaries of these themes are as follows.
The first major chapter of this thesis (Chapter 3) investigates the impact of class
imbalance on the training of FSL algorithms. Class imbalance is a common issue in
real-world datasets, where the number of samples in each class is unequal. While it
is generally understood that class imbalance harms the performance of supervised
methods, limited work examines the impact of imbalance on the FSL evaluation task
prior to this research. This thesis provides an analysis of 10 state-of-the-art meta-learning and FSL methods across different imbalance distributions and rebalancing
techniques. The results reveal that 1) some FSL methods display a natural disposition against imbalance while most other approaches produce a performance drop
by up to 17% compared to the balanced task without the appropriate mitigation;
2) contrary to popular belief, many meta-learning algorithms will not automatically
learn to balance from exposure to imbalanced training tasks; 3) classical rebalancing strategies, such as random oversampling, can still be very effective, leading to
state-of-the-art performances and should not be overlooked; 4) FSL methods are
more robust against meta-dataset imbalance than imbalance at the task-level with
a similar imbalance ratio (ρ < 20), with the effect holding even in long-tail datasets
under a larger imbalance (ρ = 65).
The second major chapter of this thesis (Chapter 4) delves into the evaluation
challenges of FSL algorithms in the open-world scenario. Typically, FSL models are
evaluated on query instances that mirror the class distribution of the support set. However, this chapter addresses the more nuanced and realistic challenge of Open-Set Few-Shot Learning (OSFSL), which involves incorporating unknown classes into
the query set. This requires the model not only to classify known classes but also
to identify outliers, reflecting the unpredictable nature of real-world evaluation settings. This scenario poses particular challenges for transductive inference methods,
which utilise the query set as an additional unlabelled set of points for enhancing the
FSL algorithm. Building on the groundwork laid by previous studies, this chapter
defines a novel transductive inference technique that leverages the InfoMax principle
to exploit the unlabelled query set. The approach is called the Enhanced Outlier
Logit (EOL) method. EOL refines class prototype representations through model
calibration and effective balancing of the inlier-outlier ratios during optimisation
within the transductive inference process. The alternations enhance pseudo-label
accuracy for the query set. The chapter also provides a comprehensive empirical evaluation demonstrating that EOL consistently surpasses traditional methods,
recording performance improvements ranging from approximately +1.3% to +6.3%
across a variety of classification and outlier detection metrics and benchmarks, even
amid inlier-outlier imbalance.
The third major chapter (Chapter 5) discusses the scalability and robustness
of FSL methods in the context of continual learning, where models must maintain
knowledge across multiple datasets. This integration poses unique challenges, as
the classical approach to continual learning (CL) does not typically address the constraints of FSL. This thesis proposes a suite of benchmarks for the hybrid setting
combining these two paradigms, where a model is trained on several sequential few-shot tasks and then tested on a validation set stemming from all those tasks. This
thesis proposes a theoretical framework for Continual Few-Shot Learning (CFSL)
with a range of flexible benchmarks to unify the evaluation criteria. A compact variant of the ImageNet dataset is introduced as part of the benchmark. The proposed
benchmarks examine several popular few-shot and continual learning methods, exposing previously unknown strengths and weaknesses of those algorithms.
The comprehensive exploration of these themes not only highlights the intricate
challenges and opportunities within FSL but also lays out practical implications and
solutions. The presented insights contribute to the development of more resilient
and adaptive FSL models, effectively bridging the gap between theoretical research
and practical deployment. Building on the solid foundation laid by this thesis,
future research is anticipated to drive the continued evolution and refinement of FSL
methodologies, ensuring they meet the rigorous demands of real-world applications
and set new benchmarks in the field.UK Engineering and Physical Sciences Research Council (Grant No. EP/S515061/1)
The restorative qualities of housing typologies : natural features at the building edge
Full text linkCity dwellers often face challenges in accessing restorative environments due
to a lack of green and blue spaces and the design and type of housing. However, the
home is a favoured setting for seeking restorative experiences. Thus, this thesis aims to
understand what features, circumstances, and experiences at the building edge
contribute to creating a restorative home. It also aims to identify and develop housing
typology based on the building edge.
To achieve these aims, the thesis adopts a mixed-methods study approach,
utilising an online survey (N=304) and participatory arts-based interviews (N=28).
However, the latter was not presented as part of the final thesis. It focuses on evaluating
Edinburgh residents' favourite indoor and outdoor spaces, as well as features such as
windows, in their current homes.
The study finds that private gardens, seating areas, lawns, and grass are the most
favourite outdoor spaces and natural/built features. These spaces exhibit restorative
qualities attributed to 1) design features and 2) the connection to the outdoors,
particularly sensory stimuli and nature. Using Latent Class Analysis (LCA), it identifies
four building edge housing typologies. It also examines their impact on residents’
general mental well-being, perceived restoration, and social and seasonal use of
outdoor space.
The research identifies preferred window types and establishes that these are
valued particularly in social spaces, kitchens and bedrooms. The restorative qualities
of window views are associated with 1) window features, 2) view attributes, and 3) the
connection to the outdoors, including sensory stimuli, nature, and people. Additionally,
the research explores the perceived benefits of observing views through residential
windows, specifically focusing on the residents’ perceived restoration from their
favourite window views and general mental well-being. The findings indicate that
individuals' satisfaction with the extent of natural views serves as a crucial predictor of
perceived restoration from favourite window views. Conversely, for residents' general
mental well-being, satisfaction with the ability to see outside whilst standing and whilst
sitting are significant predictors.
This thesis contributes to the development of design guidelines that prioritise
residents’ health and well-being through creating well-designed housing environments
that support equitable health and well-being, aligning with the third (good health and
well-being), 10th (reduced inequalities), and 11th (sustainable cities and communities)
UN Sustainable Development Goals for 2030
Access Repository Dashboard
Do you manage ROS: The Research Output Service. Heriot-Watt University Edinburgh ? Access insider analytics, issue reports and manage access to outputs from your repository in the CORE Repository Dashboard!