Deep Architecture Research Articles

Alignable kernel network

To enhance the adaptability and performance of Convolutional Neural Networks (CNN), we present an adaptable mechanism called Alignable Kernel (AliK) unit, which dynamically adjusts the receptive field (RF) dimensions of a model in response to varying stimuli. The branches of AliK unit are integrated through a novel align transformation softmax attention, incorporating prior knowledge through rank ordering constraints. The attention weightings across the branches establish the effective RF scales, leveraged by neurons in the fusion layer. This mechanism is inspired by neuroscientific observations indicating that the RF dimensions of neurons in the visual cortex vary with the stimulus, a feature often overlooked in CNN architectures. By aggregating successive AliK ensembles, we develop a deep network architecture named the Alignable Kernel Network (AliKNet). AliKNet with interdisciplinary design improves the network’s performance and interpretability by taking direct inspiration from the structure and function of human neural systems, especially the visual cortex. Empirical evaluations in the domains of image classification and semantic segmentation have demonstrated that AliKNet excels over numerous state-of-the-art architectures, achieving this without increasing model complexity. Furthermore, we demonstrate that AliKNet can identify target objects across various scales, confirming their ability to dynamically adapt their RF sizes in response to the input data.

Application of a deep-learning marker for morbidity and mortality prediction derived from retinal photographs: a cohort development and validation study

Biological ageing markers are useful to risk stratify morbidity and mortality more precisely than chronological age. In this study, we aimed to develop a novel deep-learning-based biological ageing marker (referred to as RetiPhenoAge hereafter) using retinal images and PhenoAge, a composite biomarker of phenotypic age. We used retinal photographs from the UK Biobank dataset to train a deep-learning algorithm to predict the composite score of PhenoAge. We used a deep convolutional neural network architecture with multiple layers to develop our deep-learning-based biological ageing marker, as RetiPhenoAge, with the aim of identifying patterns and features in the retina associated with variations of blood biomarkers related to renal, immune, liver functions, inflammation, and energy metabolism, and chronological age. We determined the performance of this biological ageing marker for the prediction of morbidity (cardiovascular disease and cancer events) and mortality (all-cause, cardiovascular disease, and cancer) in three independent cohorts (UK Biobank, the Singapore Epidemiology of Eye Diseases [SEED], and the Age-Related Eye Disease Study [AREDS] from the USA). We also compared the performance of RetiPhenoAge with two other known ageing biomarkers (hand grip strength and adjusted leukocyte telomere length) and one lifestyle factor (physical activity) for risk stratification of mortality and morbidity. We explored the underlying biology of RetiPhenoAge by assessing its associations with different systemic characteristics (eg, diabetes or hypertension) and blood metabolite levels. We also did a genome-wide association study to identify genetic variants associated with RetiPhenoAge, followed by expression quantitative trait loci mapping, a gene-based analysis, and a gene-set analysis. Cox proportional hazards models were used to estimate the hazard ratios (HRs) and corresponding 95% CIs for the associations between RetiPhenoAge and the different morbidity and mortality outcomes. Retinal photographs for 34 061UK Biobank participants were used to train the model, and data for 9429participants from the SEED cohort and for 3986participants from the AREDS cohort were included in the study. RetiPhenoAge was associated with all-cause mortality (HR 1·92 [95% CI 1·42-2·61]), cardiovascular disease mortality (1·97 [1·02-3·82]), cancer mortality (2·07 [1·29-3·33]), and cardiovascular disease events (1·70[1·17-2·47]), independent of PhenoAge and other possible confounders. Similar findings were found in the two independent cohorts (HR 1·67 [1·21-2·31] for cardiovascular disease mortality in SEED and 2·07 [1·10-3·92] in AREDS). RetiPhenoAge had stronger associations with mortality and morbidity than did hand grip strength, telomere length, and physical activity. We identified two genetic variants that were significantly associated with RetiPhenoAge (single nucleotide polymorphisms rs3791224 and rs8001273), and were linked to expression quantitative trait locis in various tissues, including the heart, kidneys, and the brain. Our new deep-learning-derived biological ageing marker is a robust predictor of mortality and morbidity outcomes and could be used as a novel non-invasive method to measure ageing. Singapore National Medical Research Council and Agency for Science, Technology and Research, Singapore.

Recognition of Real-Time Video Activities Using Stacked Bi-GRU with Fusion-based Deep Architecture

Recognizing and understanding human activities in real-time videos is a challenging task due to the complex nature of video data and the need for efficient and accurate analysis. This research pioneers a breakthrough in video activity recognition by introducing a robust framework leveraging the power of a stacked Bidirectional Long Short-Term Memory (Bi-LSTM) and Gated Recurrent Unit (GRU) architecture, harmonized within a fusion-based deep model. The stacked Bi-LSTM-GRU model capitalizes on its dual recurrent architecture, capturing nuanced temporal dependencies within video sequences. The fusion-based deep architecture synergizes spatial and temporal features, enabling the model to discern intricate patterns in human activities. To further enhance the discriminative power of the model, we introduce a fusion module in the proposed deep architecture. The fusion module integrates multi-modal features extracted from different levels of the network hierarchy, allowing for a more comprehensive representation of video activities. We demonstrate the efficacy of our approach through rigorous experimentation on UCF50, UCF101, and HMDB51 datasets. In experiments on the UCF50 dataset, our model achieves an accuracy of 97.01% and 95.86% on training and validation sets respectively, showcasing its proficiency in discerning activities across a diverse range of scenarios. The evaluation extends to the UCF101 dataset, where the proposed approach achieves a competitive accuracy of 97.62% and 96.93% on training and validation sets, surpassing previous benchmarks by a margin of approx 1%. Further-more, on the challenging HMDB51 dataset, the model demonstrates a robust accuracy of 89.71%and 88.88% on training and validation sets, solidifying its efficacy in intricate action recognition tasks.

Deep learning for automated encrustation detection in sewer inspection

Rapid urbanization and population growth in recent decades have placed significant pressure on urban cities to rely heavily on underground infrastructure, such as sewers and tunnels, to maintain the provision of essential services. These sewers, typically having a limited lifespan of 50 to 100 years, are prone to various forms of defects. While prior research has primarily addressed common sewer defect like crack, root intrusion, and infiltration among others, the challenge of encrustation—the formation of hard deposits within sewer systems—has received less attention. This study presents a pioneering deep-learning approach to detect encrustation in sewers by leveraging survey videos from 14 different sewers in the United Kingdom. Our work marks the first effort to develop models specifically for detecting encrustation using deep learning techniques, as previous studies have focused on other types of deposits such as settled and attached deposits. By converting the videos into sequential image frames, we subjected them to thorough analysis and several image pre-processing techniques. Our contributions include the development and comparison of different classification models using backbone CNN networks such as AlexNet, VGG16, EfficientNet, and VGG19 to classify encrustation. Notably, this study provides the first metric-based comparison of these backbone networks to identify the most effective model for encrustation detection. The results demonstrate an impressive 96 % accuracy using the deep architecture of VGG19. Beyond accuracy, this research explores the impact of data augmentation and network dropout on reducing overfitting and enhancing model performance. Additionally, we analyze the time complexities associated with training models with and without data augmentation, providing valuable insights into the efficiency of our approach.

Identifying defects and varieties of Malting Barley Kernels

This study introduces a comprehensive approach for classifying individual malting barley kernels, involving dual-sided kernel imaging, a specifically designed image processing algorithm, an optimized deep neural network architecture, and a mechanical sorting system. The proposed method achieves precise classification into multiple classes, aligning with quality standards for malting material assessment. Throughout the study, various image analysis techniques were assessed, including traditional feature engineering, established transfer learning deep neural network architectures, and our custom-designed convolutional neural network tailored for barley kernel image analysis. Comparative analysis underscores the superior performance of our network model. The study reveals that our proposed deep learning network achieves a 94% accuracy in classifying barley kernel defects and varieties, outperforming well-established transfer learning models to complex architectures that attain 93% accuracy. Additionally, it surpasses the traditional machine learning approach involving feature extraction and support vector machine classifiers, which achieve accuracy below 90% in detecting defective kernels and below 70% in varietal classification. However, we also noted the traditional approach’s advantage in morphological feature recognition. This observation guides new research toward integrating morphological feature extraction techniques with modern convolutional networks. This paper presents a deep neural network designed specifically for the analysis of cereal kernel images in two applications: defect and variety classification. It emphasizes the importance of standardizing kernel orientation and merging images from both sides of the kernel, and introduces a device for image acquisition that fulfills this need.

Accurate crowd counting for intelligent video surveillance systems

The paper presents a novel deep learning approach for crowd counting in intelligent video surveillance systems, addressing the growing need for accurate monitoring of public spaces in urban environments. The demand for precise crowd estimation arises from challenges related to security, public safety, and efficiency in urban areas, particularly during large public events. Existing crowd counting techniques, including feature-based object detection and regression-based methods, face limitations in high-density environments due to occlusions, lighting variations, and diverse human figures. To overcome these challenges, the authors propose a new deep encoder-decoder architecture based on VGG16, which incorporates hierarchical feature extraction with spatial and channel attention mechanisms. This architecture enhances the model’s ability to manage variations in crowd density, leveraging adaptive pooling and dilated convolutions to extract meaningful features from dense crowds. The model’s decoder is further refined to handle sparse and crowded scenes through separate density maps, improving its adaptability and accuracy. Evaluations of the proposed model on benchmark datasets, including Shanghai Tech and UCF CC 50, demonstrate superior performance over state-of-the-art methods, with significant improvements in mean absolute error and mean squared error metrics. The paper emphasizes the importance of addressing environmental variability and scale differences in crowded environments and shows that the proposed model is effective in both sparse and dense crowd conditions. This research contributes to the advancement of intelligent video surveillance systems by providing a more accurate and efficient method for crowd counting, with potential applications in public safety, transportation management, and urban planning.

Θ-Net: A Deep Neural Network Architecture for the Resolution Enhancement of Phase-Modulated Optical Micrographs In Silico.

Optical microscopy is widely regarded to be an indispensable tool in healthcare and manufacturing quality control processes, although its inability to resolve structures separated by a lateral distance under ~200 nm has culminated in the emergence of a new field named fluorescence nanoscopy, while this too is prone to several caveats (namely phototoxicity, interference caused by exogenous probes and cost). In this regard, we present a triplet string of concatenated O-Net ('bead') architectures (termed 'Θ-Net' in the present study) as a cost-efficient and non-invasive approach to enhancing the resolution of non-fluorescent phase-modulated optical microscopical images in silico. The quality of the afore-mentioned enhanced resolution (ER) images was compared with that obtained via other popular frameworks (such as ANNA-PALM, BSRGAN and 3D RCAN), with the Θ-Net-generated ER images depicting an increased level of detail (unlike previous DNNs). In addition, the use of cross-domain (transfer) learning to enhance the capabilities of models trained on differential interference contrast (DIC) datasets [where phasic variations are not as prominently manifested as amplitude/intensity differences in the individual pixels unlike phase-contrast microscopy (PCM)] has resulted in the Θ-Net-generated images closely approximating that of the expected (ground truth) images for both the DIC and PCM datasets. This thus demonstrates the viability of our current Θ-Net architecture in attaining highly resolved images under poor signal-to-noise ratios while eliminating the need for a priori PSF and OTF information, thereby potentially impacting several engineering fronts (particularly biomedical imaging and sensing, precision engineering and optical metrology).

Large-scale deep learning identifies the antiviral potential of PKI-179 and MTI-31 against coronaviruses

Severe Acute Respiratory Syndrome Coronavirus 2 (SARS-CoV-2) has led to the global pandemic of Coronavirus Disease (2019) (COVID-19), underscoring the urgency for effective antiviral drugs. Despite the development of different vaccination strategies, the search for specific antiviral compounds remains crucial. Here, we combine machine learning (ML) techniques with in vitro validation to efficiently identify potential antiviral compounds. We overcome the limited amount of SARS-CoV-2 data available for ML using various techniques, supplemented with data from diverse biomedical assays, which enables end-to-end training of a deep neural network architecture. We use its predictions to identify and prioritize compounds for in vitro testing. Two top-hit compounds, PKI-179 and MTI-31, originally identified as Pi3K-mTORC1/2 pathway inhibitors, exhibit significant antiviral activity against SARS-CoV-2 at low micromolar doses. Notably, both compounds outperform the well-known mTOR inhibitor rapamycin. Furthermore, PKI-179 and MTI-31 demonstrate broad-spectrum antiviral activity against SARS-CoV-2 variants of concern and other coronaviruses. In a physiologically relevant model, both compounds show antiviral effects in primary human airway epithelial (HAE) cultures derived from healthy donors cultured in an air-liquid interface (ALI). This study highlights the potential of ML combined with in vitro testing to expedite drug discovery, emphasizing the adaptability of AI-driven approaches across different viruses, thereby contributing to pandemic preparedness.

Deep Learning Methods for Chest X-Ray Imaging-Based COVID-19 Pneumonia Detection

The COVID-19 pandemic currently underway has highlighted a need for fast and accurate screening tools to help identify the disease, particularly in resource-poor settings. In this paper, a deep convolutional neural network (CNN) model is proposed for the automatic detection of COVID-19 pneumonia using chest X-ray images that helps in reducing the cost and processing time compared to traditional testing procedures. It uses several deep-structured medical image data types and improvement of the deep learning model architecture to achieve good diagnostic accuracy. Our hybrid CNN architecture integrates VGG-16 for feature extraction and ResNet-50 for pattern complexity assessment in detecting fine changes characteristic of COVID-19 pneumonia. Our dataset for this study is a collection of few thousands labelled chest X-ray images categorized in three classes, which are COVID-19, viral pneumonia and healthy cases. More advanced data preprocessing methods such as normalized, augmentation, and filtered noise has also been carried out which enhances in the model performance. We have high accuracy, recall, and good F1-score in the experimental results proving model robustness are applicable in real-world clinical scenarios. These research results speak to the importance of AI for improving diagnosis workflows, especially in fast-to-deploy large scale scenarios needed during a pandemic to cope with healthcare burdens.

Efficient physics-informed transfer learning to quantify biochemical traits of winter wheat from UAV multispectral imagery

Accurate and efficient estimation of biochemical traits, including leaf index area (LAI), leaf chlorophyll content (LCC) and canopy chlorophyll content (CCC), is crucial for crop growth monitoring in agricultural management. Recent advancements in unmanned aerial vehicle (UAV) multispectral remote sensing have enabled fast and cost-effective measurements of these traits. However, traditional statistical regression models trained on specific datasets lack scalability and transferability across practical field conditions without retraining. This study proposed an efficient physics-informed transfer learning model (PITL) for winter wheat biochemical traits estimation from UAV multispectral data. The PITL integrates the strengths of physical radiative transfer simulations and deep neural network architectures through transfer learning to improve the estimation of biochemical traits from UAV multispectral data. The PITL was tested with convolutional neural network (CNN), deep neural network (DNN), and long short-term memory (LSTM) architectures. Results indicated that PITLDNN had better accuracy than PITLCNN and PITLLSTM models in predicting LAI (R2=0.94, RMSE = 0.32 m2/m2), LCC (R2=0.81, RMSE = 5.20 μg/cm2) and CCC (R2=0.928, RMSE = 0.2 g/m2). Moreover, PITLDNN demonstrated higher capability in computational efficiency, making it suitable for processing large volumes of UAV multispectral data in crop growth monitoring applications. Furthermore, PITL's integration of radiative transfer knowledge with labeled field data yielded higher predictive accuracy compared to physically-based inversion model, pure data-driven deep neural network approaches, and hybrid models. This study highlighted the performance of PITLDNN in accurately and efficiently quantifing biochemical traits from UAV multispectral data, thereby providing timely and accurate information for guiding crop growth monitoring applications.

Analysis of Brain Age Gap across Subject Cohorts and Prediction Model Architectures.

Background: Brain age prediction from brain MRI scans and the resulting brain age gap (BAG)-the difference between predicted brain age and chronological age-is a general biomarker for a variety of neurological, psychiatric, and other diseases or disorders. Methods: This study examined the differences in BAG values derived from T1-weighted scans using five state-of-the-art deep learning model architectures previously used in the brain age literature: 2D/3D VGG, RelationNet, ResNet, and SFCN. The models were evaluated on healthy controls and cohorts with sleep apnea, diabetes, multiple sclerosis, Parkinson's disease, mild cognitive impairment, and Alzheimer's disease, employing rigorous statistical analysis, including repeated model training and linear mixed-effects models. Results: All five models consistently identified a statistically significant positive BAG for diabetes (ranging from 0.79 years with RelationNet to 2.13 years with SFCN), multiple sclerosis (2.67 years with 3D VGG to 4.24 years with 2D VGG), mild cognitive impairment (2.13 years with 2D VGG to 2.59 years with 3D VGG), and Alzheimer's dementia (5.54 years with ResNet to 6.48 years with SFCN). For Parkinson's disease, a statistically significant BAG increase was observed in all models except ResNet (1.30 years with 2D VGG to 2.59 years with 3D VGG). For sleep apnea, a statistically significant BAG increase was only detected with the SFCN model (1.59 years). Additionally, we observed a trend of decreasing BAG with increasing chronological age, which was more pronounced in diseased cohorts, particularly those with the largest BAG, such as multiple sclerosis (-0.34 to -0.2), mild cognitive impairment (-0.37 to -0.26), and Alzheimer's dementia (-0.66 to -0.47), compared to healthy controls (-0.18 to -0.1). Conclusions: Consistent with previous research, Alzheimer's dementia and multiple sclerosis exhibited the largest BAG across all models, with SFCN predicting the highest BAG overall. The negative BAG trend suggests a complex interplay of survival bias, disease progression, adaptation, and therapy that influences brain age prediction across the age spectrum.

Advancing low‐light object detection with you only look once models: An empirical study and performance evaluation

AbstractLow‐light object detection is needed for ensuring security, enabling surveillance, and enhancing safety in diverse applications, including autonomous vehicles, surveillance systems, and search and rescue operations. A comprehensive study on low‐light object detection is presented using state‐of‐the‐art you only look once (YOLO) models, including YOLOv3, YOLOv5, YOLOv6, and YOLOv8, aiming to enhance detection performance under challenging low‐light conditions. The ExDark dataset is a dataset that consists of adequate low‐light images, modified to simulate realistic low‐light scenarios, and employed for evaluation. The deep learning algorithm optimises YOLO's architecture for low‐light detection by adapting the network structure and training strategies while preserving the algorithm's integrity. The experimental results show that YOLOv8 consistently outperforms baseline models, achieving significant improvements in accuracy and robustness in low‐light scenarios. The deep learning algorithm that acquired the best score, YOLOv8s, had a mean average precision score of 0.5513. This work contributes to the field of low‐light object detection, offering promising solutions for real‐world applications like nighttime surveillance and autonomous navigation in low‐light conditions, addressing the growing demand for advanced low‐light object detection.

Emotion Recognition Using EEG Signals through the Design of a Dry Electrode Based on the Combination of Type 2 Fuzzy Sets and Deep Convolutional Graph Networks.

Emotion is an intricate cognitive state that, when identified, can serve as a crucial component of the brain-computer interface. This study examines the identification of two categories of positive and negative emotions through the development and implementation of a dry electrode electroencephalogram (EEG). To achieve this objective, a dry EEG electrode is created using the silver-copper sintering technique, which is assessed through Scanning Electron Microscope (SEM) and Energy Dispersive X-ray Analysis (EDXA) evaluations. Subsequently, a database is generated utilizing the designated electrode, which is based on the musical stimulus. The collected data are fed into an improved deep network for automatic feature selection/extraction and classification. The deep network architecture is structured by combining type 2 fuzzy sets (FT2) and deep convolutional graph networks. The fabricated electrode demonstrated superior performance, efficiency, and affordability compared to other electrodes (both wet and dry) in this study. Furthermore, the dry EEG electrode was examined in noisy environments and demonstrated robust resistance across a diverse range of Signal-To-Noise ratios (SNRs). Furthermore, the proposed model achieved a classification accuracy of 99% for distinguishing between positive and negative emotions, an improvement of approximately 2% over previous studies. The manufactured dry EEG electrode is very economical and cost-effective in terms of manufacturing costs when compared to recent studies. The proposed deep network, combined with the fabricated dry EEG electrode, can be used in real-time applications for long-term recordings that do not require gel.

Identifying Mergers in the Legacy Surveys with Few-shot Learning

Galaxy mergers exert a pivotal influence on the evolutionary trajectory of galaxies and the expansive development of cosmic structures. The primary challenge encountered in machine learning–based identification of merging galaxies arises from the scarcity of meticulously labeled data sets specifically dedicated to merging galaxies. In this paper, we propose a novel framework utilizing few-shot learning techniques to identify galaxy mergers in the Legacy Surveys. Few-shot learning enables effective classification of merging galaxies even when confronted with limited labeled training samples. We employ a deep convolutional neural network architecture trained on data sets sampled from Galaxy Zoo Decals to learn essential features and generalize to new instances. Our experimental results demonstrate the efficacy of our approach, achieving high accuracy and precision in identifying galaxy mergers with few labeled training samples. Furthermore, we investigate the impact of various factors, such as the number of training samples and network architectures, on the performance of the few-shot learning model. The proposed methodology offers a promising avenue for automating the identification of galaxy mergers in large-scale surveys, facilitating the comprehensive study of galaxy evolution and structure formation. In pursuit of identifying galaxy mergers, our methodology is applied to analyze the Data Release 9 of the Dark Energy Spectroscopic Instrument Legacy Imaging Surveys. As a result, we have unveiled an extensive catalog encompassing 648,183 galaxy merger candidates. We publicly release the catalog alongside this paper.

Tuning parameters of deep neural network training algorithms pays off: a computational study

AbstractThe paper aims to investigate the impact of the optimization algorithms on the training of deep neural networks with an eye to the interaction between the optimizer and the generalization performance. In particular, we aim to analyze the behavior of state-of-the-art optimization algorithms in relationship to their hyperparameters setting to detect robustness with respect to the choice of a certain starting point in ending on different local solutions. We conduct extensive computational experiments using nine open-source optimization algorithms to train deep Convolutional Neural Network architectures on an image multi-class classification task. Precisely, we consider several architectures by changing the number of layers and neurons per layer, to evaluate the impact of different width and depth structures on the computational optimization performance. We show that the optimizers often return different local solutions and highlight the strong correlation between the quality of the solution found and the generalization capability of the trained network. We also discuss the role of hyperparameters tuning and show how a tuned hyperparameters setting can be re-used for the same task on different problems achieving better efficiency and generalization performance than a default setting.

Dynamic functional connections analysis with spectral learning for brain disorder detection

Dynamic functional connections (dFCs), can reveal neural activities, which provides an insightful way of mining the temporal patterns within the human brain and further detecting brain disorders. However, most existing studies focus on the dFCs estimation to identify brain disorders by shallow temporal features and methods, which cannot capture the inherent temporal patterns of dFCs effectively. To address this problem, this study proposes a novel method, named dynamic functional connections analysis with spectral learning (dCSL), to explore inherently temporal patterns of dFCs and further detect the brain disorders. Concretely, dCSL includes two components, dFCs estimation module and dFCs analysis module. In the former, dFCs are estimated via the sliding window technique. In the latter, the spectral kernel mapping is first constructed by combining the Fourier transform with the non-stationary kernel. Subsequently, the spectral kernel mapping is stacked into a deep kernel network to explore higher-order temporal patterns of dFCs through spectral learning. The proposed dCSL, sharing the benefits of deep architecture and non-stationary kernel, can not only calculate the long-range relationship but also explore the higher-order temporal patterns of dFCs. To evaluate the proposed method, a set of brain disorder classification tasks are conducted on several public datasets. As a result, the proposed dCSL achieves 5% accuracy improvement compared with the widely used approaches for analyzing sequence data, 1.3% accuracy improvement compared with the state-of-the-art methods for dFCs. In addition, the discriminative brain regions are explored in the ASD detection task. The findings in this study are consistent with the clinical performance in ASD.

Training and validation of a deep learning U-net architecture general model for automated segmentation of inner ear from CT

BackgroundThe intricate three-dimensional anatomy of the inner ear presents significant challenges in diagnostic procedures and critical surgical interventions. Recent advancements in deep learning (DL), particularly convolutional neural networks (CNN), have shown promise for segmenting specific structures in medical imaging. This study aimed to train and externally validate an open-source U-net DL general model for automated segmentation of the inner ear from computed tomography (CT) scans, using quantitative and qualitative assessments.MethodsIn this multicenter study, we retrospectively collected a dataset of 271 CT scans to train an open-source U-net CNN model. An external set of 70 CT scans was used to evaluate the performance of the trained model. The model’s efficacy was quantitatively assessed using the Dice similarity coefficient (DSC) and qualitatively assessed using a 4-level Likert score. For comparative analysis, manual segmentation served as the reference standard, with assessments made on both training and validation datasets, as well as stratified analysis of normal and pathological subgroups.ResultsThe optimized model yielded a mean DSC of 0.83 and achieved a Likert score of 1 in 42% of the cases, in conjunction with a significantly reduced processing time. Nevertheless, 27% of the patients received an indeterminate Likert score of 4. Overall, the mean DSCs were notably higher in the validation dataset than in the training dataset.ConclusionThis study supports the external validation of an open-source U-net model for the automated segmentation of the inner ear from CT scans.Relevance statementThis study optimized and assessed an open-source general deep learning model for automated segmentation of the inner ear using temporal CT scans, offering perspectives for application in clinical routine. The model weights, study datasets, and baseline model are worldwide accessible.Key PointsA general open-source deep learning model was trained for CT automated inner ear segmentation.The Dice similarity coefficient was 0.83 and a Likert score of 1 was attributed to 42% of automated segmentations.The influence of scanning protocols on the model performances remains to be assessed.Graphical

Scalable spatiotemporal prediction with Bayesian neural fields

Spatiotemporal datasets, which consist of spatially-referenced time series, are ubiquitous in diverse applications, such as air pollution monitoring, disease tracking, and cloud-demand forecasting. As the scale of modern datasets increases, there is a growing need for statistical methods that are flexible enough to capture complex spatiotemporal dynamics and scalable enough to handle many observations. This article introduces the Bayesian Neural Field (BayesNF), a domain-general statistical model that infers rich spatiotemporal probability distributions for data-analysis tasks including forecasting, interpolation, and variography. BayesNF integrates a deep neural network architecture for high-capacity function estimation with hierarchical Bayesian inference for robust predictive uncertainty quantification. Evaluations against prominent baselines show that BayesNF delivers improvements on prediction problems from climate and public health data containing tens to hundreds of thousands of measurements. Accompanying the paper is an open-source software package (https://github.com/google/bayesnf) that runs on GPU and TPU accelerators through the Jax machine learning platform.

Enhanced deep potential model for fast and accurate molecular dynamics: application to the hydrated electron.

In molecular simulations, neural network force fields aim at achieving ab initio accuracy with reduced computational cost. This work introduces enhancements to the Deep Potential network architecture, integrating a message-passing framework and a new lightweight implementation with various improvements. Our model achieves accuracy on par with leading machine learning force fields and offers significant speed advantages, making it well-suited for large-scale, accuracy-sensitive systems. We also introduce a new iterative model for Wannier center prediction, allowing us to keep track of electron positions in simulations of general insulating systems. We apply our model to study the solvated electron in bulk water, an ostensibly simple system that is actually quite challenging to represent with neural networks. Our trained model is not only accurate, but can also transfer to larger systems. Our simulation confirms the cavity model, where the electron's localized state is observed to be stable. Through an extensive run, we accurately determine various structural and dynamical properties of the solvated electron.

Factorizers for distributed sparse block codes

Distributed sparse block codes (SBCs) exhibit compact representations for encoding and manipulating symbolic data structures using fixed-width vectors. One major challenge however is to disentangle, or factorize, the distributed representation of data structures into their constituent elements without having to search through all possible combinations. This factorization becomes more challenging when SBCs vectors are noisy due to perceptual uncertainty and approximations made by modern neural networks to generate the query SBCs vectors. To address these challenges, we first propose a fast and highly accurate method for factorizing a more flexible and hence generalized form of SBCs, dubbed GSBCs. Our iterative factorizer introduces a threshold-based nonlinear activation, conditional random sampling, and an ℓ ∞ -based similarity metric. Its random sampling mechanism, in combination with the search in superposition, allows us to analytically determine the expected number of decoding iterations, which matches the empirical observations up to the GSBC’s bundling capacity. Secondly, the proposed factorizer maintains a high accuracy when queried by noisy product vectors generated using deep convolutional neural networks (CNNs). This facilitates its application in replacing the large fully connected layer (FCL) in CNNs, whereby C trainable class vectors, or attribute combinations, can be implicitly represented by our factorizer having F-factor codebooks, each with C F fixed codevectors. We provide a methodology to flexibly integrate our factorizer in the classification layer of CNNs with a novel loss function. With this integration, the convolutional layers can generate a noisy product vector that our factorizer can still decode, whereby the decoded factors can have different interpretations based on downstream tasks. We demonstrate the feasibility of our method on four deep CNN architectures over CIFAR-100, ImageNet-1K, and RAVEN datasets. In all use cases, the number of parameters and operations are notably reduced compared to the FCL.