42337 research outputs found
Sort by
Real-time edge processing of neural signals with memristive technologies
No full textIntracranial brain-computer interfaces, capable of real-time neural activity decoding, present a revolutionary opportunity to improve the quality of life of individuals with dysfunction or damage to the nervous system. Despite recent advancements in neural recording, neuroprosthetic technologies still face bottlenecks in data processing and transmission. Effective neuro-prosthetic devices must deliver enhanced performance metrics including high accuracy, low-power, small size, and minimal latency to enable continuous and real-time brain interfacing. Memristive technologies are promising candidates, acting as bioelectronic links that integrate biosensing with computation for brain-inspired architectures, and operating at low power levels.
Memristive devices are two-terminal electronic components that reversibly and gradually adjust conductance in response to electrical stimuli, with their memory state depending on the thresholded integral of the input voltage. Acting as integrating sensors, they suppress noise and encode signal amplitude and frequency within their resistive state when biased with suitable preamplified neuronal signals. This behaviour similar to biological synapses offers a novel solution for processing strategies in brain-computer interfaces. A memristor-based platform for detecting action potentials (APs) -- the fundamental units of communication between neurons and a well-established indicator of brain activity -- has already demonstrated promising results in the literature.
This doctoral research proposes a memristor-based processing platform for real-time decoding of neural signals, with a focus on population-level activity rather than single-neuron action potential (AP) detection. In many clinical or assistive applications, such as state monitoring or rehabilitation relying on fine-grained single-neuron activity is not necessary. Instead, larger scale population-level dynamics provide more robust and stable biomarkers. Moreover, relying on these signals offers energy efficiency benefits due to their reduced bandwidth requirements.
Specifically, local field potentials (LFPs) were used, as they provide greater spatial coverage and temporal stability by capturing collective synaptic activity. LFPs recorded in vivo from the ventral tegmental area of awake rats performing associative memory tasks were applied to TiOx-based non-volatile memristors, significantly reducing processing power. The system achieved real-time biomarker detection with over 98% accuracy and power consumption as low as 4.14 nW per channel—up to 100× lower than comparable state-of-the-art methods, at similar accuracy levels.
This memristor-based protocol was then extended to process the envelope of multi-unit activity (eMUA), a more recently explored neural signal that also reflects population dynamics but enables earlier biomarker detection and reduced inter-channel correlation—key for real-time prosthetic control. With over 95% detection accuracy and ~9 nW power consumption, the approach was validated across different metal-oxide memristor stacks, confirming the platform-agnostic applicability of the MIS method. The integration of MIS with ultra-low-power front-end analogue circuitry showed a 30× reduction in power demand compared to majority of state of the art front-end chips, achieving sub-μW consumption and projecting up to 10× improvement over the most advanced implementations.
LFPs emerged as the most power-efficient and reliable neural source in the presented experiments, while eMUA provided a lower-latency alternative better suited to multi-channel applications. As the number of recording channels increases and monolithic integration with CMOS is optimised, this memristor-based strategy is expected to further reduce power consumption per channel while enabling the detection of increasingly complex behavioural states.
As a final experiment, given the continued prevalence of action potentials in neural signal processing, a strategy was developed to detect not amplitude-based but frequency-encoded biomarkers from action potential activity. Temporal compression was applied to reduce spiked quantity, while preserving the information needed to distinguish between high- and low-activity brain states—patterns often linked to neurological conditions such as Alzheimer’s disease or stroke. This compression was implemented using volatile metal-oxide memristors, whose intrinsic temporal filtering proved beneficial in identifying regions of high-frequency spiking activity before passing the data to a spiking neural network (SNN) for classification. Once again, neural activity was processed more efficiently and benchmarked against detection accuracy in a clinically relevant in vivo application using anaesthetised rats. These biomarkers were reliably detected using only 10% of the original data, while maintaining an SNN detection accuracy of approximately 97.5%.
Overall, this research lays the groundwork for scalable, ultra-low-power systems for chronic neural monitoring and implantable neuro-prosthetic technologies
Regression models for extreme values with random function covariates
No full textA fundamental principle of statistics of extremes is that any realistic quantification of risk requires extrapolating into a distribution’s tail—often beyond the observed extremes in a dataset.
Yet, as modern technology advances, an increasing amount of data is recorded continuously or
intermittently, and hence the question arises: how to take advantage of such data in an extreme
value framework? Motivated by this question, this thesis develops a class of novel statistical
methods that can be used for marginal and joint distributions to learn how the extreme values
may change according to a functional covariate. The first contribution consists of a functional
regression model for the tail index that can be used for assessing how the magnitude of the
extremes can change according to a random function. Another contribution of this thesis is the
development of a nonparametric regression model that can be regarded as a functional covariate
regression method, designed for situations where there is a need to assess how the extremal
dependence of a random vector can change according to a functional explanatory variable. Such
development is based on modeling a family of angular measures indexed by a random function.
The performance of the proposed methodologies is assessed via numerical studies, and financial
data is used to illustrate their application
Vibrational spectroscopy with machine learning for accurate cancer detection
No full textCancer remains a global health crisis, significantly impacting individuals and societies
worldwide. In 2020, approximately 19.3 million new cancer cases and 10 million
cancer-related deaths were reported globally. Screening and triaging are crucial in the
early detection, diagnosis, and management of cancer, targeting different stages to
improve patient outcomes. Despite being one of the leading causes of mortality, many
cancers lack effective screening methods. While conventional screening techniques are
available for some cancers, they have varying accuracy and limitations. Identifying
cancer or precancerous conditions early can significantly reduce mortality and enhance
treatment outcomes. The analysis of biofluids to detect cancer-related signals—liquid
biopsy, has garnered considerable attention over the past decade. Although promising,
many current liquid biopsies lack the sensitivity needed for early-stage cancer detection.
Raman spectroscopy (RS) is a non-destructive, real-time technique for molecular
analysis. Our study investigated the impact of optimising selected parameters and
assessed various spectral processing methods on the reliability and accuracy of spectral
analyses, and demonstrated that manual extension of the sampled volume significantly
enhanced the detection of low-concentration cancer biomolecules, improving spectral
resolution in half the measurement time compared to conventional settings.
Additionally, we examined chemical changes associated with acquired radioresistance
in HR+ and HR− breast cancer cell lines. Combining RS with machine learning, we
achieved high accuracy in distinguishing between parental cell lines and their
radioresistant phenotypes, regardless of hormonal status. The radioresistant phenotypes
exhibited similar difference spectra and formed a single cluster, suggesting common
biochemical changes during the acquisition of radioresistance. We also integrated RS
with advanced machine learning techniques for accurate cancer detection in blood
plasma, using both liquid and dried samples. Our results showed high sensitivity and
specificity in classifying stage Ia breast cancer, with an Area Under the Curve (AUC)
of 1.00. Hierarchical clustering validated the reproducibility of our results. This
research highlights the potential of combining vibrational spectroscopy with AI for
cost-effective, non-invasive, and personalised early cancer detection, emphasising the
need for standardised protocols and robust data processing techniques to facilitate
clinical translation in liquid biopsy applications
In a Widespread War of All against All, Can Ethiopia Survive the Storm?
Full text linkThis policy brief provides a narrative of Abiy Ahmed's ascent to power and his anti-establishment rhetoric, which has led the country into the current chaos and war with itself. The brief also provides background to the various wars in the country, discussing the overall alignment of forces both at regional and federal levels. The brief will also summarize the opening of the National Dialogue Commission and its achievements thus far. It concludes by highlighting some key recommendations and steps forward to end the wars in Ethiopia
Carbon Calculator for wind farms on Scottish peatlands: an evidence assessment
No full textThe Scottish Government’s Carbon Calculator for wind farms on Scottish peatlands was developed in 2008, to calculate the impact of wind farm development on peatland carbon stocks in Scotland and thereby support decision making. Electricity generation emission factors are updated annually, but no major revisions have been made to the Carbon Calculator since 2014.
The increased focus on the transition to net zero might affect the suitability of the Carbon Calculator for future use. This research conducted a detailed review of the latest spreadsheet version of the Carbon Calculator, which mirrors the web version. It provides an evidence base for future considerations and recommendations. This review has initiated further discussions and highlighted the need for ongoing engagement, which will be instrumental in any development of the Carbon Calculator
Choroidal image analysis for OCT image sequences with applications in systemic health
No full textThe retina is a light-sensitive tissue at the back of the eye and is responsible for vision. Light-sensitive photoreceptor cells in the outer retina detect light and, through a series of neuronal and vascular layers, process it into signals for the brain. The photoreceptors are perfused and maintained indirectly by the choroid and choriocapillaris, a highly vascularised layer posterior to the retina. The choroid is an extension of the central nervous system and has parallels with the renal cortex, but choroidal blood flow is four-fold higher per unit mass than the kidney and ten-fold higher than the brain. Thus, there has been growing interest in the structure and function of the choroidal circulation reflecting physiological status of systemic disease in the kidney and brain. The choroid can be imaged using optical coherence tomography (OCT), a non-invasive imaging technique which uses interferometry to capture three-dimensional, cross-sectional visualisations of ocular tissue at micron resolution. Advancements in OCT technology now permit deeper penetration and improved visualisation of the choroidal vessels. However, conventional methods of characterising and quantifying this vascular space have not kept pace with the improvements in OCT technology which visualise it, resulting in non-standardised manual or semi-automatic approaches as commonplace methods for choroidal measurement. The ability to measure anatomy consistently at micron-scale both intra- and inter-patient is paramount to capturing the inherent biological change or signal being studied – a signal which can be corrupted when exposed to human subjectivity.
In this thesis, I develop and evaluate several novel methods to analyse the choroid in OCT image sequences, with each successive method markedly improving on its predecessors. In the first instance, I develop two semi-automatic approaches for choroid region (Gaussian Process Edge Tracing, GPET) and vessel (Multi-scale Median Cut Quantisation, MMCQ) analysis, which improve on manual approaches but ultimately are biased by the end-user's biological interpretation and technical experience. As a first step to fully automatic choroid segmentation, I develop DeepGPET as a deep learning-based method for choroid region segmentation, which significantly improves on semi-automatic approaches in terms of time, reproducibility, and end-user accessibility. However, DeepGPET lacks choroidal vessel quantification and still requires manual input for generating standardised, choroid-derived measurements. Improving on this, I developed Choroidalyzer}, a fully automatic, deep learning-based, end-to-end pipeline which fully characterises the choroidal space and vessels, and automatically generates clinically meaningful and reproducible choroid-derived metrics. I provide rigorous evaluation of these four approaches, and consider their use-case and potential clinical value in three distinct applications into systemic health: OCTANE: evaluating longitudinal choroidal change and its association with renal function in transplant recipients and donors; PREVENT: investigating associations between the choroid and risk factors for developing later-life Alzheimer's disease in a mid-life cohort; D-RISCii: assessing choroidal variation and feasibility of OCT imaging in critical care. This thesis has contributed several new approaches to the research community which are all open-source and freely available, enabling consistent and reproducible measurement of the choroid. This thesis also highlights the potential role the choroid may play in reflecting pathophysiology in the kidney, brain and wider systemic health from iatrogenic shock, thus helping accelerate the nascent field of choroidal analysis in OCT image sequences
Insecure Leaders: How Elite Infighting May Facilitate the End of Iraq’s Kurdistan Region
No full textThe stability of the Kurdistan Region of Iraq (KRI) continues to deteriorate due to the power struggle between the region’s two ruling parties, the Patriotic Union of Kurdistan (PUK) and the Kurdistan Democratic Party (KDP). While they have refused to compromise and collaborate in the interest of the KRI and its citizens, the leaders of both parties have primarily relied on opportunistic tactics to weaken each other and secure short-term gains. This study attributes the sources of the ongoing PUK-KDP rivalry to leadership insecurity. Rooted in the region’s predatory system of rule, this insecurity has recently deepened due to economic and political factors, generating greater divisions between the two parties to the detriment of Kurdistan’s survival. Renewed power-sharing arrangements between the ruling parties without steps toward transforming the region’s predatory system into a democratic one will likely fail to produce long-lasting stability in the KRI
The phonetics of emphasis in Central Mount Lebanon Lebanese: acoustics, perception, articulation
No full textThis thesis is a descriptive analysis of the phonetics of emphasis in the variety of Lebanese spoken in Central Mount Lebanon (CMLL). It is both empirically and methodologically motivated, and bases its discussions around acoustic, perceptual, and articulatory data, collected from native speakers for the purposes of this thesis. The background chapters frame the variety and the phenomenon in their contexts, the data chapters dive into the experimental studies that make up the methodological core of this project, and the discussion frames the results of the different experimental approaches in relation to one another.
The acoustic study uses production data from 11 participants, recorded in Central Mount Lebanon. It finds no purely consonantal correlate to emphasis, and no statistically significant evidence towards the widely-reported F2 lowering correlate of emphasis, at vowel midpoint, across all vocalic contexts. The vowel midpoint analysis (using linear mixed models) only shows significant lowering of F2 for schwa and for long low vowels, the latter of which I report in my 2019 MScR dissertation to be categorically split into front and back low vowels with emphasis having a backing effect on the front category (/æ, ɑ/ → [ɑ] / [+EMPH]). The dynamic analysis (using univariate generalised additive mixed models) only finds a statistically significant lowering of F2 in the pre-emphatic voiceless sibilant context (_ṣ) for the low vowel, while the F2 lowering shown in other phonological contexts cannot be shown to be significant with this data. An early version of this study was published in the Journal of Semitic Studies in 2023.
The perceptual study is based on a lexical decision task given to 60 native speakers in Central Mount Lebanon. They were asked to determine whether the recording they heard matched an image associated with the plain or the emphatic member of a minimal pair. The recordings consisted of each of the original words (eg. /sīn/ versus /ṣīn/), a cross-splice of an emphatic- adjacent vowel from the word with an emphatic into the word without (eg. the /ī/ from /ṣīn/ spliced into /s_n/), and a cross-splice of the vowel in the same position from the word without the emphatic into the word with (eg. the /ī/ from /sīn/ spliced into /ṣ_n/). Using logistic regression analysis, the study finds that listeners rely on the consonant to detect emphasis in most cases, but that they instead rely on the vowel before the voiceless sibilant, in unrounded vowels after the alveolar plosives, and in unrounded vowels before the voiced alveolar plosive. As an aside, the study also shows that the emphatic-plain contrast in voiced alveolar fricatives (/ẓ/ versus /z/), while phonemic, is lexically less prominent than other such contrasts in CMLL.
The articulatory study uses ultrasound tongue imaging (UTI) data, recorded from 5 speakers of Lebanese living in Edinburgh. It explores the use of multivariate generalised additive mixed models, then dives into kinematic analyses. The data used in the study consists of CVb nonce syllables, setting the target emphatics and their unemphatic counterparts in the onset, with a voiced bilabial in the coda. It finds that there is a difference between the articulation of emphatic consonants and plain consonants. The difference in articulation occurs at the level of the pharynx, at consonant offset. The data does however suggest that the difference in articulation— this back, second constriction— occurs either further forward in the oral tract towards the uvula or velum at consonant offset, or at the pharynx earlier back in time before consonant offset, if the immediately subsequent vowel is frontal.
The discussion, grounded in theories of co-articulation, brings together the results of the three experiments, with a view to interpreting how the phonological phenomenon of emphasis operates on a phonetic level, and how the findings of three complementary experimental phonetic approaches can inform our understanding of a complex phonological phenomenon
Transmission mechanisms and persistence of the Great Depression in Italy
No full textThe first chapter – Labour Under Fascism – analyses the labour market reforms of
the Fascist regime in the late 1920s. It studies the evolution of wages, hours worked
and employment in 1930s, uncovering the behaviour of wages relative to other labour
market outcomes. It tests the hypothesis that Fascist labour market reforms increased
the degree of real wage rigidity, leading to the observed falls in employment and hours
worked. The chapter collects quantitative and qualitative evidence to determine that
wages were leading indicators of labour supply and that the 1927 reforms were associated
with a rise in estimated real wage rigidity.
The second chapter – Shielded to Shrink – provides a measure of Italy’s trade barriers
during the Great Depression, evaluating the direct losses incurred from its protectionist
stance. This chapter proposes the relationship between high trade barriers, firms’
market power and firm growth as a determining factor for the proliferation of small
businesses and the observed slowdown in industrial and aggregate total factor productivity
in 1930s Italy. Accordingly, we show that trade policies promoted industrial
concentration, stifling competition and contributing to the persistence of the Great
Depression in Italy.
The third chapter – The Battle for Credit – details the government’s rescues and
its recurrent reliance on domestic capital markets to meet its financing needs. It
shows that the price of credit on Italy’s financial markets continued to rise throughout
the 1930s, providing a framework that explains how government’s fiscal deficits could
divert resources away from the private sector. Ultimately, this chapter proposes that
crowding-out effects affected recovery by squeezing private investment and the rising
cost of credit on financial markets. We propose this mechanism to be a persistence
channel of the Great Depression in Italy
Towards robust and generalisable natural language predicate inference
Full text linkNatural language predicate inference is an important task of natural language processing,
which bears particular significance in identifying supporting material for answering
questions in the open domain. Given a premise and a hypothesis assertion, both
concerning the same entities as arguments, natural language predicate inference aims
to determine whether the hypothesis is entailed by the premise, i.e. whether the hypothesis
can be concluded from the premise.
In the task of predicate inference, directionality is a key characteristic, separating
it from symmetric semantic similarity. Directionality is the property whereby the entailment relation between a pair of assertions holds in one direction but not both. To
capture this property with an example, “John got his groceries at Tesco” entails “John
went to Tesco”; however, “John went to Tesco” does not entail “John got his groceries
at Tesco”. Directionality poses a unique challenge to predicate inference methods,
especially those not making a sufficient distinction between entailment and similarity.
Prior work has developed various approaches to natural language predicate inference.
These approaches include: 1) Entailment graphs, a type of unsupervised symbolic
methods, which are induced from predicate-argument mentions in vast natural
language corpora; 2) fine-tuned Small Language Models, a type of supervised neural
classifiers, trained on inference datasets; 3) few-shot generative interaction with Large
Language Models, another unsupervised approach whose importance has been rapidly
rising throughout the landscape of NLP and beyond.
In this thesis, we explore the development of robust and generalisable predicate inference
methods based on the above approaches. We start by generalizing entailment
graphs to other languages than English. We demonstrate that strong entailment graphs
can be built from Chinese corpora and reach strong predicate inference performance
which is comparable to their English counterparts. Additionally, between the Chinese
and English entailment graphs, we observe a mutual cross-lingual complementarity,
where a simple ensemble of mono-lingual entailment graphs elicits substantially
stronger performance than any individual graph.
On the other hand, (small) language models, with their unified representation spaces
for arbitrary text sequences, have become a common component in approaches to various
classification tasks. We hypothesized that small language models can also be
trained to perform predicate inference, and that multilingual language models would
also benefit from the cross-lingual complementarity and exhibit improved performance
when trained on multilingual inference datasets.
However, through our analysis of fine-tuned small language models, we find little
evidence of capability for the predicate inference task. Especially with directional
entailments, fine-tuned language models overfit to dataset-specific artefacts that infest
the inference datasets. On an extrinsic open-domain QA task, we observe that finetuned
LMs exhibit weak predicate inference performance in both English and Chinese,
where the performance is inferior to simple symbolic entailment graphs, despite the
latter being sparse in coverage.
With the rise of Large Language Models, which exhibit strong language understanding
performance in general and few-shot learning capability, many NLP tasks are
now considered solved by prompting LLMs with few-shot examples to generate the
predictions. Since the few-shot in-context learning paradigm is independent of training
datasets and the artefacts in them, we are able to use this setup to investigate the
LLMs’ predicate inference capability. Through controlled experiments, we observe
positive results overall, where few-shot LLMs yield non-trivial performance; on the
other hand, we also observe worrying statistical biases in LLM answer generation. We
identify prominent types of such biases, including an attestation bias and a relative
frequency bias. We argue that these biases must be alleviated before LLMs can be
trusted to perform robust natural language predicate inference in context.
Overall, we develop and critically compare approaches to predicate inference, and
advocate for a hybrid approach between efficient symbolic entailment graphs and versatile
neural LLMs
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!