Sparse Datasets Research Articles

GrapHiC: An integrative graph based approach for imputing missing Hi-C reads.

Hi-C experiments allow researchers to study and understand the 3D genome organization and its regulatory function. Unfortunately, sequencing costs and technical constraints severely restrict access to high-quality Hi-C data for many cell types. Existing frameworks rely on a sparse Hi-C dataset or cheaper-to-acquire ChIP-seq data to predict Hi-C contact maps with high read coverage. However, these methods fail to generalize to sparse or cross-cell-type inputs because they do not account for the contributions of epigenomic features or the impact of the structural neighborhood in predicting Hi-C reads. We propose GrapHiC, which combines Hi-C and ChIP-seq in a graph representation, allowing more accurate embedding of structural and epigenomic features. Each node represents a binned genomic region, and we assign edge weights using the observed Hi-C reads. Additionally, we embed ChIP-seq and relative positional information as node attributes, allowing our representation to capture structural neighborhoods and the contributions of proteins and their modifications for predicting Hi-C reads. We show that GrapHiC generalizes better than the current state-of-the-art on cross-cell-type settings and sparse Hi-C inputs. Moreover, we can utilize our framework to impute Hi-C reads even when no Hi-C contact map is available, thus making high-quality Hi-C data accessible for many cell types. Availability: https://github.com/rsinghlab/GrapHiC.

Extended Features Based Random Vector Functional Link Network for Classification Problem

Random vector functional link (RVFL) network has been successfully employed in diverse domains such as computer vision and machine learning, due to its universal approximation capability. Recently, the shallow RVFL architecture has been extended to deep architectures. In deep architectures, multiple hidden layers are stacked for extracting informative features from the original feature space. Therefore, having rich features, deep models are very successful compared to shallow models. In this article, we propose an extended feature RVFL (efRVFL) model that is trained over extended feature space generated analytically from the original feature space. The proposed efRVFL model has three types of features, i.e., original features, supervised randomized (newly generated) features, and unsupervised randomized features, in its feature matrix. The proposed efRVFL model with additional features has capability to capture nonlinear hidden relationships within the dataset. The proposed efRVFL model is an unstable classifier, and thus, its performance can be improved further via ensemble learning. Ensemble models are stable and accurate and have better generalization performance than single models. Therefore, we also propose an ensemble of extended feature RVFL (en-efRVFL) model. Each base model of en-efRVFL is trained over different feature spaces so that more accurate and diverse base models can be generated. The outcome of the base models is integrated via average voting scheme. Empirical evaluation over <inline-formula> <tex-math notation="LaTeX">$46$</tex-math> </inline-formula> UCI classification datasets demonstrates that the proposed efRVFL and en-efRVFL models have better performance than RVFL and other given deep models. Furthermore, the experimental results over <inline-formula> <tex-math notation="LaTeX">$12$</tex-math> </inline-formula> sparse datasets show that the proposed en-efRVFL model has a winning performance among several deep feedforward neural networks (FNNs).

Assessment of Domestic Water Resources for Sustainable Utilization Using Geospatial Techniques. The Case of Pune City, India

Water resources have played a crucial role in the growth and development of human civilisation. Increasing demands of water resources to meet the needs of growing population have led to tremendous pressure on water resources. Water resources conservation and management need relevant information regarding the quality and quantity of water bodies, as well as the related driving factors responsible for the deterioration and depletion of water resources. In literature, conventional methods are limited to point locations that have sparse datasets. However, with the advent of geospatial techniques, it has become very easy to explore digital information that can quickly support extensive data analysis and interpretation on a larger scale. The aim of this study is to assess water resources such as water supply, distribution and coverage using geospatial techniques, and also to identify the water stress zones whilst forecasting the water demand in the future, which is indispensable for sustainable urban planning. The results show that water supply and duration per capita are unequal in the study area. The water demand of Pune city has continuously increased from 508 MLD in 2001 to 857 MLD in 2022, and will maintain this trend up to 1421 MLD by 2031. The composite water stress map indicates that Ghole Road, Nagar Road, Dhankawadi and Hadapsar wards have a high level of water stress, whereas Sangamwadi, Kothrud, Dhole Patil Road, Bhawani Peth, Kasba Vishrambagh and Sahakranagar experience low water stress. The values of water scarcity index in the study area have significantly changed, i.e., from 0.68 in 2001, to 0.54 in 2011 and to 0.64 in 2022, which indicates that Pune city falls into the serious water shortage category. This index is expected to be around 1.05 in 2031, demonstrating that Pune will be subjected to major water deficiency, a condition which is clearly reflected in the composite water stress map.

Pragmatic degradation learning for scene text image super-resolution with data-training strategy

Super-resolution of scene text images represents a formidable computational problem, marred by a myriad of intricate challenges. This paper focuses on the specific hurdles that have impeded significant advancements in this domain, and introduces the Higher-Order Degradation-Based Super-Resolution Network (HDSN) as a novel solution to address these intricate issues. The challenges in super-resolving scene text images are manifold. Firstly, the semantic ambiguity inherent to text in natural scenes often leads to degraded results, as standard super-resolution techniques struggle to preserve meaningful textual content. Additionally, the uncertainty surrounding font variability exacerbates this issue, as different fonts require distinct treatment for optimal super-resolution. Furthermore, scene text images often exhibit long trailing shadows, artifacts, and strong noise, rendering conventional methods inadequate in producing satisfactory results. To tackle these intricate challenges, we propose a pragmatic higher-order degradation modeling process. This process takes into account the nuanced characteristics of scene text images, including the diverse forms of noise such as Gaussian, Poisson, speckle, and JPEG compression noise, as well as varying levels of blurring. By meticulously considering these real-world scenarios, our approach significantly enhances the robustness and adaptability of super-resolution for scene text images. In addition to addressing these challenges, we recognize the issues arising from sparse datasets and the lack of corresponding paired images for training. To surmount this limitation, we introduce a text image pre-training strategy, which proves to be highly effective in improving recognition accuracy. The experimental results on TextZoom affirm the effectiveness of our approach, demonstrating substantial improvements over existing methods. Notably, our HDSN achieves average recognition rates of 67.2% on ASTER, 63.2% on MORAN, and 58.0% on CRNN, surpassing the performance of available approaches. Our source code is available at https://github.com/syyang2022/HDSN.

A Comprehensive Evaluation of Generalizability of Deep Learning-Based Hi-C Resolution Improvement Methods.

Hi-C is a widely used technique to study the 3D organization of the genome. Due to its high sequencing cost, most of the generated datasets are of a coarse resolution, which makes it impractical to study finer chromatin features such as Topologically Associating Domains (TADs) and chromatin loops. Multiple deep learning-based methods have recently been proposed to increase the resolution of these datasets by imputing Hi-C reads (typically called upscaling). However, the existing works evaluate these methods on either synthetically downsampled datasets, or a small subset of experimentally generated sparse Hi-C datasets, making it hard to establish their generalizability in the real-world use case. We present our framework-Hi-CY-that compares existing Hi-C resolution upscaling methods on seven experimentally generated low-resolution Hi-C datasets belonging to various levels of read sparsities originating from three cell lines on a comprehensive set of evaluation metrics. Hi-CY also includes four downstream analysis tasks, such as TAD and chromatin loops recall, to provide a thorough report on the generalizability of these methods. We observe that existing deep learning methods fail to generalize to experimentally generated sparse Hi-C datasets, showing a performance reduction of up to 57%. As a potential solution, we find that retraining deep learning-based methods with experimentally generated Hi-C datasets improves performance by up to 31%. More importantly, Hi-CY shows that even with retraining, the existing deep learning-based methods struggle to recover biological features such as chromatin loops and TADs when provided with sparse Hi-C datasets. Our study, through the Hi-CY framework, highlights the need for rigorous evaluation in the future. We identify specific avenues for improvements in the current deep learning-based Hi-C upscaling methods, including but not limited to using experimentally generated datasets for training.

Improving Landslide Prediction: Innovative Modeling and Evaluation of Landslide Scenario with Knowledge Graph Embedding

The increasing frequency and magnitude of landslides underscore the growing importance of landslide prediction in light of factors like climate change. Traditional methods, including physics-based methods and empirical methods, are beset by high costs and a reliance on expert knowledge. With the advancement of remote sensing and machine learning, data-driven methods have emerged as the mainstream in landslide prediction. Despite their strong generalization capabilities and efficiency, data-driven methods suffer from the loss of semantic information during training due to their reliance on a ‘sequence’ modeling method for landslide scenarios, which impacts their predictive accuracy. An innovative method for landslide prediction is proposed in this paper. In this paper, we propose an innovative landslide prediction method. This method designs the NADE ontology as the schema layer and constructs the data layer of the knowledge graph, utilizing tile lists, landslide inventory, and environmental data to enhance the representation of complex landslide scenarios. Furthermore, the transformation of the landslide prediction task into a link prediction task is carried out, and a knowledge graph embedding model is trained to achieve landslide predictions. Experimental results demonstrate that the method improves the F1 score by 5% in scenarios with complete datasets and 17% in scenarios with sparse datasets compared to data-driven methods. Additionally, the application of the knowledge graph embedding model is utilized to generate susceptibility maps, and an analysis of the effectiveness of entity embeddings is conducted, highlighting the potential of knowledge graph embeddings in disaster management.

Identifying accurate artefact morphological ranges using optimal linear estimation: Method validation, case studies, and code

A fundamental goal of archaeologists is to infer the behaviour of past humans from the attributes of the artefacts they left behind. The archaeological record is, however, fragmented and often provides a partial record of the total artefacts produced by a given population. In turn, there is potential for population-level morphometric data, and therefore behavioural inferences, to be biased relative to the trends realised in the past. This includes morphological range data which are important for identifying similarities and differences between artefact groups, and for contextualising artefacts relative to external variables such as human anatomy, ecology, climate and chronology. Here, we investigate whether optimal linear estimation (OLE) modelling can be used to accurately identify the upper and lower limits of artefact morphological ranges (including those represented by sparse datasets). First, we test whether OLE reliably identifies morphological ranges using randomly sampled subsets of ‘known and complete’ replica artefact assemblages. Using morphometric data from lithic, ceramic, and metal archaeological case studies, we then identify how much further the upper and lower form limits of these artefact types would have been in the past, relative to the ranges evidenced by excavated (i.e., known partial) records. Validation tests demonstrate OLE to be capable of providing broadly accurate estimates for the true morphological range of artefact assemblages. Estimate accuracy increases relative to the percentage of the total assemblage used and the method is shown to function well using as few as five records (k) from an assemblage. The case studies reveal how OLE can overhaul or reinforce our understanding of artefact morphological ranges. In some instances, it is clear that the archaeological record provides a highly accurate representation of artefact morphological ranges and the overlap between artefact groups. For others, it is demonstrated that our understanding of the extreme artefact forms produced by past people is likely inaccurate

High-resolution reconstruction of turbulent flames from sparse data with physics-informed neural networks

Accurate and detailed data are vital for fundamental understanding of turbulent combustion. However, studies of turbulent combustion often suffer from measurement sparsity or high simulation cost. In the present work, for the first time, a physics-informed neural networks (PINNs) framework is established for three-dimensional high-resolution reconstruction of turbulent combustion with synthetic sparse data. The performance of the PINNs is evaluated on two different configurations of turbulent flames without and with a mean shear, including freely propagating planar premixed combustion and slot-jet premixed combustion. The reconstructed fields of velocity, temperature, and species mass fractions are compared with the high-fidelity direct numerical simulation (DNS) data, and both qualitative and quantitative analyses are performed for the model performance evaluation. The results show that by constraining with the residuals of the governing equations and limited information of sparse data, the proposed PINNs can recover the majority of flow and flame structures in a physics-informed way, even when noise has been added to the sparse data. It is noted that the effectiveness of the models can be influenced by various factors such as the size of the sparse dataset, noise levels, and complexity of turbulence/flame interactions. These factors should be carefully evaluated and addressed to ensure reliable predictions. Overall, this study highlights the potential of PINNs for data assimilation and provide new insights for the development of physics-informed methods in combustion research.Novelty and significanceIn the present work, a physics-informed neural networks (PINNs) framework for turbulent combustion has been established and the possibility of PINNs for high-resolution reconstruction of various field data in turbulent combustion was explored, which is the first of its kind. The velocity, temperature and species mass fractions were simultaneously reconstructed from limited pointwise sparse data, which were compared with the high-fidelity direct numerical simulation data with good agreements. The proposed PINNs have the capability to recover the flow and flame structures for turbulent flames without and with mean shear layers. Various factors such as the size of the sparse dataset, noise level, and complexity of turbulence/flame interactions on the model performance were also assessed. The study highlights the potential of PINNs for data assimilation and provide new insights for the development of physics-informed methods in combustion research.

Communities in C. elegans connectome through the prism of non-backtracking walks

The fundamental relationship between the mesoscopic structure of neuronal circuits and organismic functions they subserve is one of the major challenges in contemporary neuroscience. Formation of structurally connected modules of neurons enacts the conversion from single-cell firing to large-scale behaviour of an organism, highlighting the importance of their accurate profiling in the data. While connectomes are typically characterized by significant sparsity of neuronal connections, recent advances in network theory and machine learning have revealed fundamental limitations of traditionally used community detection approaches in cases where the network is sparse. Here we studied the optimal community structure in the structural connectome of Caenorhabditis elegans, for which we exploited a non-conventional approach that is based on non-backtracking random walks, virtually eliminating the sparsity issue. In full agreement with the previous asymptotic results, we demonstrated that non-backtracking walks resolve the ground truth annotation into clusters on stochastic block models (SBM) with the size and density of the connectome better than the spectral methods related to simple random walks. Based on the cluster detectability threshold, we determined that the optimal number of modules in a recently mapped connectome of C. elegans is 10, which precisely corresponds to the number of isolated eigenvalues in the spectrum of the non-backtracking flow matrix. The discovered communities have a clear interpretation in terms of their functional role, which allows one to discern three structural compartments in the worm: the Worm Brain (WB), the Worm Movement Controller (WMC), and the Worm Information Flow Connector (WIFC). Broadly, our work provides a robust network-based framework to reveal mesoscopic structures in sparse connectomic datasets, paving way to further investigation of connectome mechanisms for different functions.

Unprecedented distribution data for Joshua trees (Yucca brevifolia and Y. jaegeriana) reveal contemporary climate associations of a Mojave Desert icon

IntroductionForecasting range shifts in response to climate change requires accurate species distribution models (SDMs), particularly at the margins of species' ranges. However, most studies producing SDMs rely on sparse species occurrence datasets from herbarium records and public databases, along with random pseudoabsences. While environmental covariates used to fit SDMS are increasingly precise due to satellite data, the availability of species occurrence records is still a large source of bias in model predictions. We developed distribution models for hybridizing sister species of western and eastern Joshua trees (Yucca brevifolia and Y. jaegeriana, respectively), iconic Mojave Desert species that are threatened by climate change and habitat loss.MethodsWe conducted an intensive visual grid search of online satellite imagery for 672,043 0.25 km2 grid cells to identify the two species' presences and absences on the landscape with exceptional resolution, and field validated 29,050 cells in 15,001 km of driving. We used the resulting presence/absence data to train SDMs for each Joshua tree species, revealing the contemporary environmental gradients (during the past 40 years) with greatest influence on the current distribution of adult trees.ResultsWhile the environments occupied by Y. brevifolia and Y. jaegeriana were similar in total aridity, they differed with respect to seasonal precipitation and temperature ranges, suggesting the two species may have differing responses to climate change. Moreover, the species showed differing potential to occupy each other's geographic ranges: modeled potential habitat for Y. jaegeriana extends throughout the range of Y. brevifolia, while potential habitat for Y. brevifolia is not well represented within the range of Y. jaegeriana.DiscussionBy reproducing the current range of the Joshua trees with high fidelity, our dataset can serve as a baseline for future research, monitoring, and management of this species, including an increased understanding of dynamics at the trailing and leading margins of the species' ranges and potential for climate refugia.

Multi-hop path reasoning over sparse temporal knowledge graphs based on path completion and reward shaping

Multi-hop path reasoning plays a crucial role in temporal knowledge graph reasoning, aiming to infer deep complex relationships and obtain interpretable reasoning results. Previous reasoning models primarily focus on designing temporal knowledge graphs with rich paths between entities but struggle to handle path sparsity, known as sparse temporal knowledge graphs. Sparse temporal knowledge graphs store only essential knowledge information, resulting in missing connection paths between certain entities and encountering issues of sparse rewards and information scarcity. To tackle these challenges, this paper proposes a multi-hop path reasoning model over sparse temporal knowledge graph based on path completion and reward shaping (STKGR-PR). STKGR-PR dynamically completes missing paths using a temporal embedding model to alleviate path sparsity. To tackle the issue of sparse rewards caused by reduced hit rates, we propose semantic composition-based reasoning path embeddings derived from relation embeddings and timestamp embeddings for reward shaping. Considering the limited information contained in sparse temporal knowledge graphs, we incorporate time vectors into the embedding and multi-hop path reasoning models to enhance the accuracy of reasoning paths and results. Experimental results conducted on six sparse temporal datasets sampled from ICEWS14 and ICEWS515–05 demonstrate that STKGR-PR outperforms state-of-the-art multi-hop path reasoning models over temporal knowledge graphs across all evaluation metrics, while ensuring interpretability.

Assessing the potential of synthetic and ex situ airborne laser scanning and ground plot data to train forest biomass models

Abstract Airborne laser scanning data are increasingly used to predict forest biomass over large areas. Biomass information cannot be derived directly from airborne laser scanning data; therefore, field measurements of forest plots are required to build regression models. We tested whether simulated laser scanning data of virtual forest plots could be used to train biomass models and thereby reduce the amount of field measurements required. We compared the performance of models that were trained with (i) simulated data only, (ii) a combination of simulated and real data, (iii) real data collected from different study sites, and (iv) real data collected from the same study site the model was applied to. We additionally investigated whether using a subset of the simulated data instead of using all simulated data improved model performance. The best matching subset of the simulated data was sampled by selecting the simulated forest plot with the highest correlation of the return height distribution profile for each real forest plot. For comparison, a randomly selected subset was evaluated. Models were tested on four forest sites located in Poland, the Czech Republic, and Canada. Model performance was assessed by root mean squared error (RMSE), squared Pearson correlation coefficient (r$^{2}$), and mean error (ME) of observed and predicted biomass. We found that models trained solely with simulated data did not achieve the accuracy of models trained with real data (RMSE increase of 52–122 %, r$^{2}$ decrease of 4–18 %). However, model performance improved when only a subset of the simulated data was used (RMSE increase of 21–118 %, r$^{2}$ decrease of 5–14 % compared to the real data model), albeit differences in model performance when using the best matching subset compared to using a randomly selected subset were small. Using simulated data for model training always resulted in a strong underprediction of biomass. Extending sparse real training datasets with simulated data decreased RMSE and increased r$^{2}$, as long as no more than 12–346 real training samples were available, depending on the study site. For three of the four study sites, models trained with real data collected from other sites outperformed models trained with simulated data and RMSE and r$^{2}$ were similar to models trained with data from the respective sites. Our results indicate that simulated data cannot yet replace real data but they can be helpful in some sites to extend training datasets when only a limited amount of real data is available.

Solving the imbalanced data issue: automatic urgency detection for instructor assistance in MOOC discussion forums

In MOOCs, identifying urgent comments on discussion forums is an ongoing challenge. Whilst urgent comments require immediate reactions from instructors, to improve interaction with their learners, and potentially reducing drop-out rates—the task is difficult, as truly urgent comments are rare. From a data analytics perspective, this represents a highly unbalanced (sparse) dataset. Here, we aim to automate the urgent comments identification process, based on fine-grained learner modelling—to be used for automatic recommendations to instructors. To showcase and compare these models, we apply them to the first gold standard dataset for Urgent iNstructor InTErvention (UNITE), which we created by labelling FutureLearn MOOC data. We implement both benchmark shallow classifiers and deep learning. Importantly, we not only compare, for the first time for the unbalanced problem, several data balancing techniques, comprising text augmentation, text augmentation with undersampling, and undersampling, but also propose several new pipelines for combining different augmenters for text augmentation. Results show that models with undersampling can predict most urgent cases; and 3X augmentation + undersampling usually attains the best performance. We additionally validate the best models via a generic benchmark dataset (Stanford). As a case study, we showcase how the naïve Bayes with count vector can adaptively support instructors in answering learner questions/comments, potentially saving time or increasing efficiency in supporting learners. Finally, we show that the errors from the classifier mirrors the disagreements between annotators. Thus, our proposed algorithms perform at least as well as a ‘super-diligent’ human instructor (with the time to consider all comments).

Performance Analysis of Sketching Methods

One of the problems in data mining or machine learning is the high-dimensional dataset. MinHash can generate sketches of sparse datasets efficiently reducing the dimension to a few thousand. These sketches, then, can be used for different machine-learning applications. It takes O(kd) computations to generate k hash values for a data point with d non-zeros. The Sketch of a data point is the vector of k hash values. Weighted Minwise Hashing is another method to generate a sketch of size k in O(kp/d) computations. Here, p is the size of the universal set. Optimal Densification is the most efficient and accurate method as it can generate k hash values in mere O(d + k) computations. In this paper, we investigate the performance of Optimal Densification, Weighted Minwise Hashing, and Vanilla MinHash by performing two experiments. Firstly, we investigate Jaccard similarity estimation accuracy on six different synthetic datasets. Then, we perform one nearest neighbor classification (1NN) of four real datasets. Optimal Densification outperforms both Weighted Minwise Hashing and Vanilla MinHash in terms of accuracy and time taken to generate the sketches.

ScFed: federated learning for cell type classification with scRNA-seq.

The advent of single-cell RNA sequencing (scRNA-seq) has revolutionized our understanding of cellular heterogeneity and complexity in biological tissues. However, the nature of large, sparse scRNA-seq datasets and privacy regulations present challenges for efficient cell identification. Federated learning provides a solution, allowing efficient and private data use. Here, we introduce scFed, a unified federated learning framework that allows for benchmarking of four classification algorithms without violating data privacy, including single-cell-specific and general-purpose classifiers. We evaluated scFed using eight publicly available scRNA-seq datasets with diverse sizes, species and technologies, assessing its performance via intra-dataset and inter-dataset experimental setups. We find that scFed performs well on a variety of datasets with competitive accuracy to centralized models. Though Transformer-based model excels in centralized training, its performance slightly lags behind single-cell-specific model within the scFed framework, coupled with a notable time complexity concern. Our study not only helps select suitable cell identification methods but also highlights federated learning's potential for privacy-preserving, collaborative biomedical research.

Sparse convolutional neural network for high-resolution skull shape completion and shape super-resolution

Traditional convolutional neural network (CNN) methods rely on dense tensors, which makes them suboptimal for spatially sparse data. In this paper, we propose a CNN model based on sparse tensors for efficient processing of high-resolution shapes represented as binary voxel occupancy grids. In contrast to a dense CNN that takes the entire voxel grid as input, a sparse CNN processes only on the non-empty voxels, thus reducing the memory and computation overhead caused by the sparse input data. We evaluate our method on two clinically relevant skull reconstruction tasks: (1) given a defective skull, reconstruct the complete skull (i.e., skull shape completion), and (2) given a coarse skull, reconstruct a high-resolution skull with fine geometric details (shape super-resolution). Our method outperforms its dense CNN-based counterparts in the skull reconstruction task quantitatively and qualitatively, while requiring substantially less memory for training and inference. We observed that, on the 3D skull data, the overall memory consumption of the sparse CNN grows approximately linearly during inference with respect to the image resolutions. During training, the memory usage remains clearly below increases in image resolution—an times 8 increase in voxel number leads to less than times 4 increase in memory requirements. Our study demonstrates the effectiveness of using a sparse CNN for skull reconstruction tasks, and our findings can be applied to other spatially sparse problems. We prove this by additional experimental results on other sparse medical datasets, like the aorta and the heart. Project page at https://github.com/Jianningli/SparseCNN.

Stochastic Gradient Descent for matrix completion: Hybrid parallelization on shared- and distributed-memory systems

The purpose of this study is to investigate the hybrid parallelization of the Stochastic Gradient Descent (SGD) algorithm for solving the matrix completion problem on a high-performance computing platform. We propose a hybrid parallel decentralized SGD framework with asynchronous inter-process communication and a novel flexible partitioning scheme to attain scalability up to hundreds of processors. We utilize Message Passing Interface (MPI) for inter-node communication and POSIX threads for intra-node parallelism. We tested our method by using different real-world benchmark datasets. Experimental results on a hybrid parallel architecture showed that, compared to the state-of-the-art, the proposed algorithm achieves 6× higher throughput on sparse datasets, while it achieves comparable throughput on relatively dense datasets.

Pixel-Level Degradation for Text Image Super-Resolution and Recognition

In the realm of image reconstruction, deep learning-based super-resolution (SR) has established itself as a prevalent technique, particularly in the domain of text image restoration. This study aims to address notable deficiencies in existing research, including constraints imposed by restricted datasets and challenges related to model generalization. Specifically, the goal is to enhance the super-resolution network’s reconstruction of scene text image super-resolution and utilize the generated degenerate dataset to alleviate issues associated with poor generalization due to the sparse scene text image super-resolution dataset. The methodology employed begins with the degradation of images from the MJSynth dataset, using a stochastic degradation process to create eight distinct degraded versions. Subsequently, a blank image is constructed, preserving identical dimensions to the low-resolution image, with each pixel sourced randomly from the corresponding points across the eight degraded images. Following several iterations of training via Finetune, the LR-HR method is applied to the TextZoom dataset. The pivotal metric for assessment is optical character recognition (OCR) accuracy, recognized for its fundamental role in gauging the pragmatic effectiveness of this approach. The experimental findings reveal a notable enhancement in OCR accuracy when compared to the TBSRN model, yielding improvements of 2.4%, 2.3%, and 4.8% on the TextZoom dataset. This innovative approach, founded on pixel-level degradation, not only exhibits commendable generalization capabilities but also demonstrates resilience in confronting the intricate challenges inherent to text image super-resolution.

Language, artificial education, and future-making in indigenous language education

ABSTRACT This paper examines how language-based artificial intelligence is envisaged to imagine new futures for indigenous languages. It draws on the visions, programmes, and plans of six language initiatives that are developing language technology for often-marginalised indigenous, tribal, and minority (ITM) languages, such as Gondi, Maithili, Rajasthani and Mundari, in India. We note three distinct discourses: (1) technological optimism in utilising these new opportunities by claiming space for otherwise-marginalised languages, (2) the imperative for collaborative and collective work in order to address sparse datasets, and (3) the need to negotiate the contested nature of imagining a new collective future. This paper argues that indigenous language technology is not just a technical project but a contested process of subverting linguistic hierarchy through the ‘active presencing’ of these languages. Overall, the paper emphasizes the need for a nuanced approach that recognizes the interplay between technology, language education, and broader social and political factors.

Assessing optimal methods for transferring machine learning models to low-volume and imbalanced clinical datasets: experiences from predicting outcomes of Danish trauma patients.

Accurately predicting patient outcomes is crucial for improving healthcare delivery, but large-scale risk prediction models are often developed and tested on specific datasets where clinical parameters and outcomes may not fully reflect local clinical settings. Where this is the case, whether to opt for de-novo training of prediction models on local datasets, direct porting of externally trained models, or a transfer learning approach is not well studied, and constitutes the focus of this study. Using the clinical challenge of predicting mortality and hospital length of stay on a Danish trauma dataset, we hypothesized that a transfer learning approach of models trained on large external datasets would provide optimal prediction results compared to de-novo training on sparse but local datasets or directly porting externally trained models. Using an external dataset of trauma patients from the US Trauma Quality Improvement Program (TQIP) and a local dataset aggregated from the Danish Trauma Database (DTD) enriched with Electronic Health Record data, we tested a range of model-level approaches focused on predicting trauma mortality and hospital length of stay on DTD data. Modeling approaches included de-novo training of models on DTD data, direct porting of models trained on TQIP data to the DTD, and a transfer learning approach by training a model on TQIP data with subsequent transfer and retraining on DTD data. Furthermore, data-level approaches, including mixed dataset training and methods countering imbalanced outcomes (e.g., low mortality rates), were also tested. Using a neural network trained on a mixed dataset consisting of a subset of TQIP and DTD, with class weighting and transfer learning (retraining on DTD), we achieved excellent results in predicting mortality, with a ROC-AUC of 0.988 and an F2-score of 0.866. The best-performing models for predicting long-term hospitalization were trained only on local data, achieving an ROC-AUC of 0.890 and an F1-score of 0.897, although only marginally better than alternative approaches. Our results suggest that when assessing the optimal modeling approach, it is important to have domain knowledge of how incidence rates and workflows compare between hospital systems and datasets where models are trained. Including data from other health-care systems is particularly beneficial when outcomes are suffering from class imbalance and low incidence. Scenarios where outcomes are not directly comparable are best addressed through either de-novo local training or a transfer learning approach.