42337 research outputs found
Sort by
Effects of context on semantic representations and mechanisms in humans and language models
Get PDFThe rapid and incremental nature of language processing is a central challenge for human cognition. Understanding how this challenge is met has resulted in a broad range of work focused on answering questions such as elucidating which mechanisms support processing and exploring which computational models serve as good approximates for language processing. Typically, this work has focused on semantic processing across single words or sentences. However, we now know that context plays an important role in how the semantic system processes linguistic inputs. In this thesis, I investigate the influence of context on semantic representations and mechanisms in humans and language models. As much of the literature has focused on single sentence contexts, I investigate context at both wider and narrower scales. The first branch of studies focus on this wider scale by investigating the impact of discourse coherence on predictive processing in both humans and Large Language Models (LLMs). The second branch of studies focus on the narrower scale by exploring the influence of context on pre-trained word embeddings in a perceptual property prediction task for both nouns and adjective–noun phrases. In addition, I investigated how a neural network encodes perceptual features in conceptual combinations. In the first branch of work, I found that human’s lexical–semantic predictions are sensitive to discourse coherence, but especially so when semantic violations are present. From modelling, I found that LLMs are similarly sensitive to the relationship between context and a target sentence. This is in addition to coherence effects and their interaction with predictability, which suggests that the benefit of a highly coherent context extends beyond just lowering linguistic surprisal. In the second branch of work, I found reasonable performance for the perceptual prediction of the shape of a concept from word embeddings, but lower performance for the brightness of a concept. This was not impacted by contextual prompting for noun representations, though I did find a limited impact of context when predicting the brightness of adjective–noun pairs. This has implications for the interpretability of representations derived from language models and for debates on embodiment within human conceptual processing. In the final study, I found that neural networks can flexibly encode the modulation of conceptual features when nouns are modified with scalar adjectives. They do this by first learning to generate predictions based on the adjective, and then acquiring knowledge of how the adjective modulates particular nouns. In sum, this thesis adds greater depth to our understanding of how context influences language in humans and machines
Mesoscale particle-based modelling of active nematic liquid crystals
Get PDFActive matter --- materials with energy injection at local scales --- has developed rapidly in the past few decades, with applications ranging from the macroscopic scale of crowds and animal flocks, down to the mesoscale of bacteria colonies and active gels, and finally to the microscopic scale of sub-cellular fluids. Active fluids exhibit feedback loops that can drive or mitigate activity. For example, in the quintessential experimental active-nematic system of microtubule bundles interlinked with kinesin molecular motors activity has a sigmoidal dependence on the density of ATP fuel. Likewise, bundles of contractile nematic actin-myosin systems can form into dense jammed asters, locking in myosin molecular motors and jamming the actin, and bacteria have been shown to perform quorum sensing by inhibiting their motion upon receipt of a chemical secreted by other bacteria. Despite the rapid development of active matter theory and experiment, there are still significant gaps in our understanding: Notably, there are limited simulation methods suitable for studies at the mesoscale. For instance, despite many numerical studies investigating the behaviour of colloidal particles in athermal baths, there has been little work investigating how oriented active flows, such as active-nematics, affect the dynamics of colloids.
To address these gaps, this thesis presents a particle-based mesoscale simulation method known as Multi-Particle Collision Dynamics (MPCD) for simulating active fluctuating nematohydrodynamics. It extends an existing algorithm for nematic fluids in MPCD to produce an active-nematic MCPD method (AN-MPCD) through the introduction of a local force dipole. Despite its simplicity, AN-MPCD reproduces key quantities of active-nematic turbulence, such as spontaneous flows and the continuous creation/annihilation of topological defects. This simple model exhibits pronounced density fluctuations, typical of active particle models. By de-coupling the magnitude of activity from the local density, or applying a sigmoidal modulation function with respect to the local density, we show that density fluctuations are strongly mitigated while key scalings of active-nematic turbulence persist. Analysis of the density-induced pressure gradients reveal that activity modulation suppresses the effect of density fluctuations on solutes in active fluids, at the expense of fluctuations in the active force. Finally, we employ the modulated AN-MPCD method to study anchored passive colloids in active nematics. Homeotropic colloids possess non-monotonic effective diffusion, giving rise to a critical activity for enhanced diffusivity. When the colloidal radius is comparable to the active nematic length scale, active turbulence causes a colloidal companion defect to unbind from the colloid, leading to a non-zero topological charge on the colloid-companion complex. This non-zero charge encourages oppositely charged defects to approach the complex, indirectly propelling the complex through the fluid and enhancing the effective diffusion.
The development of the AN-MPCD method opens up a wide range of possibilities for the study of solutes immersed in active solvents, including passive colloids, polymers, and porous media, but can be extended further to novel systems of passive clusters in active fluids. It also opens the door to future studies of how the formulation of activity and spatial modulation can affect bulk active turbulence
Can public libraries still be public? A case study of the introduction of maker education as a new educational activity in the makerspaces of the Amsterdam Public Library
Get PDFThis thesis presents a case study on the introduction of maker education as a new educational activity in the makerspaces of the Amsterdam Public Library. In these newly established makerspaces, children can use crafts and modern technologies like 3D printers and laser cutters to create artefacts. This research aims to better understand the changing educational and public roles of public libraries in the Netherlands and discuss what introducing new activities means for their future development.
The research used a case study design, including a historical and socio-political context analysis and an empirical study. The empirical study involved ten in-depth interviews with maker space coaches to illuminate their educational perceptions and ambitions. The interviews were recorded, transcribed, and analysed using thematic analysis.
In this study, Habermas’ (1991) public sphere concept was used to examine the public role of libraries in conjunction with the triadic model of educational aims (Aspin & Chapman, 2001, 2012; Biesta, 2006) and Regmi's (2017) educational processes of meaning-making, opinion-formation and intersubjective deliberation. From this perspective, public libraries can be perceived as institutions that contribute to the common good based on democratic values of equality and social justice and are responsive to diverse societal needs. Additionally, public libraries function as intermediate institutions in deliberative democracy, thematising societal issues via collections, exhibitions, and events to facilitate opinion-formation processes in civil society. Finally, they can be seen as public institutions from their educational role if emancipatory and democratic aims are part of their educational mission.
The historical analysis showed that libraries have played a vital public role in various eras. However, recent socio-political changes indicate that the public role of libraries is declining. Library policies focus on lifelong learning, citizenship, and active social participation, emphasising individual skill development as solutions for structural, social, and economic problems. In addition, the recent introduction of accountability systems and output measurement on preset policy goals poses the risk of top-down control of public institutions. These systems impede professional judgment and responsiveness to the context-specific needs of library users and the democratic participation of the public in library policy.
Analysing the narratives of makerspace coaches revealed that coaches do not adhere to the agenda of lifelong learning and active and responsible citizenship. Instead, they address broader educational aims, including ‘agency and voice’. Their educational role involves thematising relevant societal issues closely related to children's daily experiences. Moreover, the coaches have a social and political awareness that reflects the public values of the library. They describe an educational practice not standardised or regulated by narrow, measurable objectives. Their approach varied depending on the location, the children's backgrounds and the coaches’ interests.
The research shows that if educational activities are developed outside the policy of schooling and without interference from accountability systems, there is still an opportunity to resist the neoliberal agenda of individual skill development and reconnect to the library's history as an educational and public institution. The research indicates that, under specific circumstances, small-scale bottom-up initiatives involving professionals, partners, and social movements can develop. These initiatives enrich educational discourses and potentially strengthen the library as a public educational good.
Nevertheless, the future of these makerspaces is ambiguous. Further institutionalisation may lead to a loss of the qualities observed in this research. Future research should explore the sustainability of makerspaces within the Amsterdam Public Library. Furthermore, new case studies should determine if the case in Amsterdam is unique or if more initiatives reimagine the library as a public educational good. Additionally, systematic research is needed to understand the conditions and support systems fostering librarians as normative educational professionals and education as a context-based practice
From Substate Governance to Constitution-building at the Centre: A View from Somalia
Get PDFSince the effective collapse of central authority in Somalia in 1991, local systems have provided crucial platforms for governance. Local systems of governance continue to be relevant alongside efforts to rebuild the Somali state, such as attempts to revive formal pre-1991 local government institutions.
This case study uses interviews and desk research to explore the ideas and practices underlying local governance frameworks, and the challenges facing official local government formation processes in Galmudug and Hirshabelle, two federal member states in Somalia. The overall objective is to understand whether—and if so how—local values and institutions can offer insights into ongoing efforts to develop practical and acceptable constitutional frameworks at the federal member state and federal levels
Exploring adolescents' sexual health literacy with adolescents and sexuality educators in Macau: a participatory action research study
No full textBACKGROUND:
Adolescents often report dissatisfaction with sexuality education in family and school settings, which is typically adult-led, abstinence-focused and framed by heteronormativity. Consequently, adolescents frequently seek alternative information sources, such as the internet and peers, though these landscapes present their own barriers, including reliability and normative biases. Sexual health literacy is therefore essential for navigating this complex information landscape, challenging dominant sexual discourses, addressing unequal power dynamics and promoting individual and collective sexual well-being. However, research on adolescent sexual health literacy often provides conflicting evidence and lacks meaningful adolescent engagement. This study aims to fill these gaps by exploring adolescent sexual health literacy through collaboration with adolescents and sexuality educators in Macau.
METHODS:
This two-phase online participatory action research study used photovoice with 16 adolescents and four sexuality educators across four online sessions. Analysis was conducted collaboratively with adolescents and sexuality educators, using Wang and Burris’ (1997) photovoice analysis and Gale and colleagues’ (2013) framework analysis.
RESULTS:
Three overarching themes were generated, with each consisting of two sub- themes, five key dimensions and two suggestions. Based on the data, adolescents often internalised and actively enforced adult-imposed abstinence, heterosexuality and gender norms, resulting in the prevailing sexual silences in the families and schools. Despite the tightly woven power relations, adolescents gained access to more open perspectives online, often from Western sources, enabling them to critically reflect on local norms. However, they faced challenges with online information, including limited access to reliable local information, the idealisation of Western sexuality, reinforcement of gender stereotypes and the perpetuation of peer sexual cyberbullying. Adolescents and educators recommended listening to adolescents’ voices, providing more reliable, positive and inclusive information across both online and offline platforms and fostering a more inclusive environment in both digital and offline spaces.
DISCUSSION:
This study is among the first to my knowledge, to use participatory action research to explore adolescent sexual health literacy in Macau. By engaging both adolescents and sexuality educators in dialogue, it provides new knowledge and valuable insights into existing challenges, priorities and potential actions to strengthen adolescent sexual health literacy in Macau and beyond
Generating environments and pre-training agents for efficient reinforcement learning
Get PDFReinforcement learning provides a compellingly universal approach for learning to achieve an objective, specified by a reward function, by trial-and-error interaction with an environment. While this approach has the versatility to be applied to almost any objective in any environment, it can be inhibitively inefficient. Humans are able to learn to achieve new objectives and improve their capabilities via reinforcement learning over human timescales by building on prior knowledge and capabilities. However, a large proportion of the reinforcement literature considers the traditional problem of learning to perform a task tabula-rasa. In this thesis, we aim to improve the efficiency of reinforcement learning, including both the sample efficiency and computational efficiency, by incorporating environment understanding and knowledge of prior behaviours via more information-dense supervised learning objectives.
In the first half of the thesis, we aim to acquire knowledge about the environment that can be leveraged for reinforcement learning. We begin by considering how to optimally combine an agent’s partial observations into a unified representation of an environment. We introduce a novel approach that can more effectively integrate partial information into a single representation than other self-supervised approaches. We next develop this general idea of learning environment representations into a diffusion-based approach for learning a full generative model of an environment. An agent can then perform model-based reinforcement learning by interacting with its environment model rather than the true environment, thereby reducing the environment dependency, and improving the sample efficiency of reinforcement learning. We demonstrate that our diffusion-based approach can more effectively capture visual details compared to related world modelling approaches, leading to greater performance and sample efficiency. However, this doesn’t reduce the computational cost of reinforcement learning, and in fact increases it, due to the additional cost of environment modelling.
In the second half of the thesis, we therefore aim to reduce the computational cost of tabula-rasa reinforcement learning by incorporating imitation learning on prior behaviours to provide an initial behaviour that can be efficiently improved with model-free reinforcement learning. We begin by considering the offline-only case from proprioceptive states with a clear objective, and demonstrate our proposed value-based approach leads to improved performance and computational efficiency over imitation and reinforcement learning approaches in isolation. We then extend this general idea to the more general offline-to-online case from visual observations without a well-defined reward function. The training procedure proposed is analogous to that used for modern large language models, providing many exciting directions for future research. We conclude by considering the future directions of generative world models and generalist agents
The design of radiation hydrodynamical simulations for the helium reionisation epoch
No full textThe reionisation of the intergalactic medium (IGM) is a critical phase transition
in the history of our Universe. The start of the reionisation of the IGM signifies
the emergence of energetic objects in the Universe, such as the earliest stars,
galaxies and quasi-stellar objects (QSOs). Observational data from HI Lymanalpha
forest spectra and HeII Lyman-alpha forest spectra during the epoch of
hydrogen reionisation (EoR) and the epoch of helium reionisation (EoHeR) (HeII
reionisation) provide valuable insights into the large-scale structure and evolution
of galaxies and QSOs. Interpreting these observations requires prior knowledge
of the reionisation processes of the IGM, including the temperature, ionisation
fractions and patchiness of the IGM at different epochs, which can only be
estimated through numerical simulations. However, modelling the reionisation
processes evolving at relativistic speeds, driven by super-luminous sources like
QSOs, poses significant challenges for existing simulation methods.
In this thesis, I improve and examine several ray-tracing algorithms to identify
the key numerical factors for simulating the reionisation processes driven by
QSOs. Central to this work is enhancing the ray-tracing radiative transfer
module of the cosmological hydrodynamics code ENZO. My contributions include
the implementation of a probabilistic absorption method, an adaptive timestep
scheme accounting for HeII ionisation, and crucial corrections to the source
code. These modifications address critical issues such as non-conservative photon
counts, inaccuracies originating from a non-cosmological adaptive timestep
scheme, and the superluminal propagation of ionisation fronts resulting from
the infinite speed of light approximation in simulations. Consequently, these
enhancements improve the reliability and accuracy of ENZO in modelling ionisation
processes in the proximity zone of QSOs. This work also underscores the necessity
of incorporating time-dependent calculations in radiation transfer simulations.
To further address issues associated with the finite speed of light, I develop a
v
novel three-dimensional time-dependent ray-tracing code, 3DPhRay, specifically
designed to accurately account for the finite propagation speed of radiation
fields. By propagating rays at the physical speed of light, 3DPhRay correctly
solves the time-dependent radiative transfer equation in a static space. The code
employs an adaptive ray-tracing scheme, a probabilistic absorption method, and
a geometric correction method to ensure photon conservation and accuracy in
ionisation calculations. Additionally, 3DPhRay is written in C++ and utilises
OpenMP for parallel computation, thereby optimising performance on modern
multicore processors.
Finally, to illustrate the impact of the propagation speed of radiation fields in
simulations, I apply 3DPhRay to simulate the 21-cm signature produced by highmass
X-ray binaries (HMXBs) in a massive early galaxy during Cosmic Dawn. By
incorporating the time delay effect due to the finite speed of light, the simulations
predict that such galaxies can additionally heat their surrounding IGM, creating
a warm halo with a distinctive 21-cm signature. Detecting these signals would
confirm the hypothesis that X-rays produced in early galaxies contribute to the
IGM heating during Cosmic Dawn.
The research presented in this thesis constitutes an advance in the accurate
simulation of the reionisation processes of the IGM, particularly during the
EoHeR. The enhancements to ENZO provide a reliable cosmological radiation
hydrodynamical framework for a wide range of research. The development of
3DPhRay offers a robust numerical scheme for accurately modelling phenomena
that evolve at relativistic speed, such as QSO-driven IGM reionisation. These
tools are crucial for deepening our understanding of the thermal history of the
IGM in the Universe. A precise description of the history of the evolution of
the IGM will also facilitate the interpretation of observational data and so allow
refinement of constraints on the formation of the large-scale structure in the
Universe
High dynamic range single photon avalanche diode pixels for time of flight imaging
No full textLight detection and ranging (LiDAR) is a distance determination method involving illumination of an object in front of the camera with a pulsed light source and measuring the time needed for the light to return after the reflection from the target. LiDARs are used extensively for the creation of high-resolution maps, monitoring of the natural environment and ranging for military applications. They are also increasingly adopted in the automotive sector where they can provide indispensable situational awareness for autonomous vehicles. There is a variety of LiDAR implementations, including scanned solutions where the field of view is swept point-by-point by the laser, and flash LiDARs that illuminate the entire scene at once. Both of these architectures require a suitable light detector, while the latter also necessitates the use of a 2D pixel array. Flash systems are advantageous in the applications where the camera or objects are moving, as they avoid motion artifacts.
Single photon avalanche diodes (SPADs) are PN junctions reversely biased above their breakdown voltage and thus sensitive to single quanta of light due to a high (“infinite”) device gain. They signal photon detection with a picosecond resolution and are therefore especially useful in building of time of flight (ToF) systems. However, they pose a few challenges in their implementation in outdoor LiDARs. Most notably, due to the very nature of SPADs reacting to single impinging photons, their dynamic range (DR) can be severely limited when the photon arrival rate is increasing. At the same time, imaging in high solar background conditions is safety-critical for automotive LiDAR applications where inability to detect or accurately measure the distance to an object in front of the moving car often poses a threat to human life. Moreover, abundant retroreflective surfaces on the road signs further increase the chance of saturating the pixel. In recent years there have been many efforts in increasing the dynamic range of SPAD-based pixels, involving highly-optimized devices and different schemes of quenching, recharging and processing the photon counts, out of which many are implemented in 3-D-stacked processes. Despite these efforts, there is still a need for simple, compact and high dynamic range SPAD pixels suitable for applications in high-resolution sensors.
This thesis researches the techniques of high dynamic range SPAD-based pixel design for LiDAR and other ToF systems. First, the response of the SPAD-based ToF detector is modelled and simulated in different environmental conditions, relating to the outdoor use of LiDAR with high solar background. The analytical model of the SPAD device with the frontend circuits is developed in order to study the behaviour of an RC-coupled SPAD in near-paralysis regime and to simulate the response of the SPAD pixel to different illumination levels while varying a SPAD output pulse detection threshold. It is noted that early detection of the avalanche by the pixel frontend is useful in picking up small-amplitude voltage excursions which occur close to SPAD saturation. Leveraging this, a set of novel SPAD pixel frontends is proposed, utilizing compact, near-threshold inverter-based circuit and other architectures. The circuits are simulated and benchmarked for use in high-pixel-count arrays using Monte Carlo methods by means of electronic design automation (EDA) tools, including simulation program with integrated circuit emphasis (SPICE) for full-custom analogue design. In order to further evaluate the envisaged pixel frontends, a test chip is designed and fabricated in STMicroelectronics’ 3D40 stacked technology. Test structures include standalone SPAD frontends, small SPAD array and a directly-accessible diode for device characterization.
The proposed frontends are characterized and confirmed to deliver significant improvements in dynamic range and maximum count rate. They are compared against a standard foundry frontend in state-of-the-art 3-D-stacked back-side illuminated 65 nm/40 nm process, achieving up to 26 dB higher DR and two-times higher maximum count rate. The latter is improved by even larger factor when compared against recently published state-of-the-art 3-D-stacked SPAD pixels from other manufacturers. Photoresponse curve measurements show that the peak irradiance is moved by 20 dB towards higher light intensities for the best performing structures. Additionally, developed circuits are implemented mostly using small and low-power thin-oxide 40 nm transistors, which makes them compact and easily scalable to high-resolution pixel arrays, providing and enabling factor for the development of future, high-performance SPAD-based LiDAR systems
High-throughput characterisation of suspensions
No full textSuspensions are ubiquitous in the modern world, including in food, medicine, and nature. Therefore, beyond academic study, they are of significant interest to industry and to society more widely. One of the most important characteristics of these suspensions is the particle size distribution, which impacts suspension stability, reaction rates, sensory properties, and product safety. A wide array of techniques exist for measuring size, but challenges arise when the distribution is broad and/or multimodal. In addition, these techniques are often slow, expensive, or both. A relatively new alternative is Differential Dynamic Microscopy (DDM), which can be cheap to implement and provides similar ensemble-averaged data to the widely used technique of dynamic light scattering (DLS).
This thesis provides a comprehensive guide to particle sizing with DDM, covering experimental considerations, data analysis methods, and interpretation of results. We begin with a review of common sizing techniques and key considerations, highlighting challenges introduced by multimodal suspensions. We then collate practical guidelines for sizing with DDM, many of which are not yet present in the literature, and provide methods to reduce the impact of user choices and improve reproducibility. A theoretical framework is established for understanding signal generation in DDM of suspensions, which shows that signal contribution of each particle scales with the sixth power of its radius, and the square of its form factor. We demonstrate experimentally, using a well-characterised model system, that this provides specific advantages over DLS when working with multimodal suspensions, and study the points at which analysis breaks down. Finally, to showcase the technique on a real-world material, DDM is used to simultaneously size protein micelles and fat globules in milk, a challenging task which has led to significant debate in literature. Our approach is validated with intensive application of Cryo-FIB-SEM nanotomography, and this serves as a practical case-study of the strengths and limitations of DDM
Estimating fabric energy efficiency using current Energy Performance Certificate metrics
No full textThis study investigates the use of Energy Performance Certificate (EPC) metrics as a predictor of the new fabric rating metric for energy efficiency of domestic properties in Scotland
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!