Gustavus Adolphus College Collections
Not a member yet
92910 research outputs found
Sort by
Online Task-Free Continual Learning via Expansible Vision Transformer
No full textVision Transformers (ViTs) have lately shown remarkable data representation capabilities leading to state-of-the-art results in several vision and language learning tasks. Given its powerful representation ability, some recent studies have explored the ViT in continual learning by employing the dynamic expansion mechanism. However, these methods rely on the task information and therefore can not deal with a more realistic scenario, namely the Task-Agnostic Continual Learning (TACL). Unlike these ViT-based continual learning methods, this paper addresses TACL by proposing the Lifelong Expansible Vision Transformer (LEViT) model, which dynamically increases the model’s capacity to deal with changes in the underlying probability distribution of the data representations learnt during continual learning. LEViT is implemented by an ensemble of transformers, each enabled with a multi-head attention mechanism and a linear classifier. We propose a new dynamic expansion mechanism which incrementally increases the capacity of LEViT without requiring task labels, by evaluating the statistical similarity between the joint distribution modeled by all previously learned components and the probabilistic representation of incoming data samples. The proposed expansion mechanism ensures the diversity of learnt knowledge by the components of LEViT. In addition, we introduce the Dynamic Knowledge Fusion (DKF) approach, aiming to explore the ViT feature representation ability for knowledge transfer. Specifically, we view all previously learnt components as an evolved knowledge base which provides prior knowledge for future learning. The proposed LEViT, when compared to the existing ViT-based methods, does not require any task information and can reuse previously learned representations to promote future task learning
Multi-layer process control in selective laser melting: a reinforcement learning approach
No full textPowder bed fusion (PBF) is an original additive manufacturing technique for creating 3D parts layer-by-layer. While there are numerous benefits to this process, the complex undergoing physical phenomena are challenging to analytically model and interpret. Hence, integrated and control-oriented 3D models are lacking in the current literature. As a result, the state of the art in process control for the powder bed fusion (PBF) process is not as advanced as in other manufacturing processes. Reinforcement learning is a machine learning, data-driven mathematical and computational framework that can be used for process control while addressing this challenge (lack of control-oriented models) effectively. Its flexible formulation and its trial-and-error nature make reinforcement learning suitable for processes where the model is intricate or even unknown. The focus of this research work is selective laser melting, which is a laser-based PBF process. For the first time in the literature we demonstrate the benefits of a reinforcement learning process control framework for multiple layers (complete 3D parts) and we highlight the importance of stability during training. The presented case studies confirm the effectiveness of the proposed control framework, directly addressing heat accumulation issues while demonstrating effective overall process control, hence opening up opportunities for further research and impact in this area
Forgotten fibre waste: Mycoremediation and recycling of used absorbent hygiene products
No full textBeyond clothing, end-of-life technical textiles and nonwoven products are an additional source of fibre-based waste and environmental impact. Absorbent Hygiene Products (AHPs), i.e. single use, disposable diapers (nappies), adult incontinence and menstrual products, play an important role in supporting the personal hygiene and wellbeing of millions of people worldwide, but their disposal presents considerable waste management and environmental challenges due to their biological contamination, as well as mixed fibre and polymer composition. Despite high rates of consumption, used AHPs remain one of the hardest waste streams to recycle, and most are incinerated or landfilled. Internationally, very little used AHP recycling infrastructure exists, and generating high-value outputs from such waste is highly challenging, mainly due to its multifaceted nature. This review evaluates the potential for an alternative biotechnological approach to recycling based on mycoremediation and biocatalysis of used AHPs (containing cellulose, superabsorbent polymers and synthetic polymers) harnessing fungi to valorise the cellulosic and plastic components of the waste. We focus on the synergistic integration of mycoremediation and precision fermentation techniques as part of a biorefinery model to yield valuable material outputs from used AHPs, such as industrial chemicals and fibre-forming biodegradable polymers for industrial applications, as a basis for new circular economies
Adhesive layer formation and its dual role in tribological performance and surface integrity of Ti-6Al-4V: Implications for the machining process
No full textThe poor machinability of Ti-6Al-4V (Ti64), characterized by adhesive and abrasive wear, low thermal conductivity, and high chemical reactivity, continues to hinder efficient manufacturing. Among these challenges, adhesive layer formation on tool flank faces remains poorly understood despite its critical influence on tool degradation and workpiece surface integrity. To address this, this study investigates the tribological behavior of WC/Co-Ti64 pin-on-disc sliding contacts under dry and minimum quantity lubrication (MQL) conditions through both experimental and numerical approaches. Experimental results show that thick, stable, and intact adhesive layers transferred from Ti64 discs was formed on WC/Co pin surfaces under dry and low MQL flowrate conditions. These layers are associated with reduced friction coefficients and lower disc wear but simultaneously contribute to compromised surface integrity. Comparative boundary element method (BEM) simulations with 316 L stainless steel reveal that the lower elastic modulus of Ti64 adhesive layers significantly reduces nominal contact pressure and subsurface von Mises stress, lowering friction coefficients and enhancing mechanical stability of adhesive layer. However, the accompanying increase in surface roughness intensifies local stress concentrations and result in thicker work-hardened layers on Ti64 disc, which align well with BEM simulation results. Conversely, high MQL flowrate inhibited adhesive layer formation, leading to higher friction and wear but producing smoother surfaces and thinner work-hardened layer. The findings offer new mechanistic insights into complex interplay between adhesive layer, lubrication and surface topography, and present the first direct evidence of the dual role of adhesive layer: reducing friction and tool-side wear but compromising workpiece surface integrity
Efficient network compression via gradient-score aware pruning
No full textConvolutional neural networks (CNNs) have demonstrated significant achievements in the field of computer vision, yet their high computational demands restrict practical applications. Current pruning methods seek to mitigate this issue, which however often rely on heuristic manual approaches, encountering challenges in maintaining both significant model compression and accuracy. To address the above issues, a fast neural architecture search pruning (FNP) technique is proposed in this paper. Firstly, an importance matrix (IM) based preprocessing stage efficiently removes redundant structures by considering both weight importance and computational complexity, providing a compact baseline for subsequent pruning. Secondly, we adapt fast genetic algorithms (FGA) to identify optimally pruned model configurations. Furthermore, to accelerate the search process, we utilize a zero-shot learning approach to estimate model performance with the score of the frame (SoF), which is a gradient-based score. Compared with state-of-the-art (SOTA) pruning techniques, FNP demonstrates superior performance in terms of search duration and compression ratio. On the CIFAR-10 dataset, our method removes 95.24 % of the parameters in VGG-16 while achieving a 0.72 % accuracy improvement compared with the baseline. On the ImageNet dataset, we prune 68.98 % of the parameters in ResNet-50 and obtain a 1.2 % accuracy improvement compared with state-of-the-art (SOTA) approaches, while reducing the search time by 98.94 %. The code is available at https://github.com/aqiu1222/FNP.gi
Which trees matter most? The role of private garden trees and woodland cover for 3–30-300 success in seven English cities
No full textThe 3–30–300 rule is a tool to evaluate access to trees and greenspaces and is gaining popularity in Europe but not yet in the UK. We calculate a 3–30–300 score per building to measure success at the rule in the local authority areas of seven English cities, examining how overall canopy cover and where the canopy is situated (e.g. woodland, street, private garden) influence performance. We find that a maximum of 2.1 % of buildings in the locations studied meet all three rules. Land use analysis indicates that increasing the density of trees in private gardens and increasing woodland cover are the most important factors for improving performance at the 3-tree and 30 % components in UK neighbourhoods. These recommendations should be applied to UK urban areas to improve overall performance at the 3–30–300 rule and increase access to trees and their benefits. We also explore how sensitive the results of the 3–30–300 analysis are to methodological choices by comparing results of network and line-of-sight analyses to simple buffers for the 3-tree and 300 m components of the rule, finding that more simple methods result in higher 3–30–300 scores and therefore suggest better 3–30–300 performance
Uneven development and the geographies of energy transition in Mozambique
No full textIn Mozambique, sustainable energy access is an increasing priority for a diverse range of actors seeking to improve livelihoods and stimulate economic development—particularly in rural areas where energy infrastructure remains limited. Drawing on field research conducted as part of a comparative three-year project examining the potential of community energy systems to foster inclusive, just, and clean energy transitions in Southern and East Africa, this paper develops a critical, policy-relevant and geographically grounded analysis of Mozambique’s energy transitions, which are unfolding across multiple fronts. Our analysis addresses both the move away from conventional or ‘traditional’ energy sources to ‘modern’ energy services, and the shift toward renewable energy technologies. We argue that while Mozambique has taken important steps toward a cleaner energy future there remain significant constraints to progress and that it is crucial to consider the advancement of renewable energy in relation to the country’s embedded resource and extractive geographies that shape the directions, possibilities, and spatial dynamics of transition. We examine the broader policy environment, focusing on the state’s energy transition strategy and its implications for energy justice, spatial inequality, and economic opportunity. Particular attention is given to the role and potential of decentralized, off-grid energy systems, emphasizing the need for greater community participation in both policy design and implementation. Finally, we develop a political economy framework to analyse the influence of state institutions, international donors, and private capital in shaping Mozambique’s energy transitions, and assess their impacts on energy poverty and the goal of equitable, sustainable energy access
Building evidence regarding nature-based solutions indicators and their implications for policy – the case of air quality
No full textAir pollution is one of world's largest planetary health risk factors. Nature-based solutions (NBS) have been key in integrating air quality indicators into the urban green planning and public health discourse. Despite important contributions, approaches that include multidimensional indicators into research, planning and policies are still limited. National standards for some types of air pollutants are missing, with little evidence for a threshold for health effects. To respond to these gaps, we provide an overview and guidance for air pollution indicators, using three case studies in Europe and Latin America. We discuss the importance of context, specific pollutants and vulnerable groups and suggest new approaches at finer scales. Our findings also point out that knowledge of pollutants uptake in edible plants can give a hint to potential exposure risks for humans. Our lessons learned target specific policies and are organized into three main ideas: (a) multidimensional indicators and their implications for NBS and policy; (b) plants as biological indicators and as schools’ subjects and (c) the integration of the co-benefits to manage air quality
Shakedown limit analysis for heavy-haul railway tracks
No full textThe lower-bound shakedown theorem provides a useful framework for evaluating the long-term stability of structures subjected to cyclic loading by defining both the shakedown limit and critical depth. However, its application in freight railway engineering remains relatively limited. To overcome this gap, shakedown theory has been integrated into the design of heavy-haul railway trackbed systems, enabling assessment of substructure stability under repeated loading. The stress distributions along the longitudinal and transverse axes of the sub-ballast surface, induced by a four-axle loading pattern, were quantified and validated through Gaussian curve fitting. Additionally, a methodology based on the Mohr–Coulomb yield criterion was developed to estimate the shakedown limit of the subgrade, employing the corresponding shakedown axle load as the primary evaluation index. Parametric analyses examined the effects of three key design parameters: the internal friction angle of the sub-ballast, the elastic modulus of the engineered subgrade, and the thickness ratio between the sub-ballast and engineered subgrade. Findings consistently showed that increases in these parameters lead to higher shakedown axle loads. Among them, the internal friction angle of the sub-ballast has the most pronounced influence, whereas the thickness ratio plays a relatively minor role. For example, elevating the internal friction angle from 25° to 40° produces a significant 47.5% rise in the shakedown axle load, highlighting its pivotal contribution to enhancing the substructure's resilience against cyclic loads
A platform for CRISPRi-seq in Streptomyces albidoflavus
No full textStreptomyces produce a multitude of secondary metabolites, which have been exploited in drug discovery campaigns for more than three-quarters of a century. Our understanding of microbial physiology has been revolutionized by genome sequencing and large-scale functional studies. Technology for genome-wide investigations in Streptomyces species, however, has lagged behind that for other bacterial systems, hindering exploitation of unprecedented quantities of genomic data. Here, we develop a platform for en masse clustered regularly interspaced short palindromic repeats interference sequencing (CRISPRi-seq) for Streptomyces spp. By performing CRISPRi-seq with 2,160 unique sgRNAs targeting all operons (432 operons) encoding membrane transporters (629 genes) representing 1.1Mb of the 6.8Mb genome for S. albidoflavus, combined with hit validation, we discovered that only a small proportion (13 of 432 operons, 25 kb) contribute positively to fitness. Our work provides both a first-in-class platform for high-throughput functional genomics and a generalized blueprint for en masse screens in Streptomyces species