Training Deep Neural Networks Research Articles

Deep-Interior: A new pathway to interior tomographic image reconstruction via a weighted backprojection and deep learning.

In recent years, deep learning strategies have been combined with either the filtered backprojection or iterative methods or the direct projection-to-image by deep learning only to reconstruct images. Some of these methods can be applied to address the interior reconstruction problems for centered regions of interest (ROIs) with fixed sizes. Developing a method to enable interior tomography with arbitrarily located ROIs with nearly arbitrary ROI sizes inside a scanning field of view (FOV) remains an openquestion. To develop a new pathway to enable interior tomographic reconstruction for arbitrarily located ROIs with arbitrary sizes using a single trained deep neural network model. The method consists of two steps. First, an analytical weighted backprojection reconstruction algorithm was developed to perform domain transform from divergent fan-beam projection data to an intermediate image feature space, , for an arbitrary size ROI at an arbitrary location inside the FOV. Second, a supervised learning technique was developed to train a deep neural network architecture to perform deconvolution to obtain the true image from the new feature space . This two-step method is referred to as Deep-Interior for convenience. Both numerical simulations and experimental studies were performed to validate the proposed Deep-Interior method. The results showed that ROIs as small as a diameter of 5 cm could be accurately reconstructed (similarity index 0.985 ± 0.018 on internal testing data and 0.940 ± 0.025 on external testing data) at arbitrary locations within an imaging object covering a wide variety of anatomical structures of different body parts. Besides, ROIs of arbitrary size can be reconstructed by stitching small ROIs without additionaltraining. The developed Deep-Interior framework can enable interior tomographic reconstruction from divergent fan-beam projections for short-scan and super-short-scan acquisitions for small ROIs (with a diameter larger than 5 cm) at an arbitrary location inside the scanning FOV with high quantitative reconstruction accuracy.

A scalable pipeline to create synthetic datasets from functional–structural plant models for deep learning

Abstract In plant science, it is an established method to obtain structural parameters of crops using image analysis. In recent years, deep learning techniques have improved the underlying processes significantly. However, since data acquisition is time and resource consuming, reliable training data are currently limited. To overcome this bottleneck, synthetic data are a promising option for not only enabling a higher order of correctness by offering more training data but also for validation of results. However, the creation of synthetic data is complex and requires extensive knowledge in Computer Graphics, Visualization and High-Performance Computing. We address this by introducing Synavis, a framework that allows users to train networks on real-time generated data. We created a pipeline that integrates realistic plant structures, simulated by the functional–structural plant model framework CPlantBox, into the game engine Unreal Engine. For this purpose, we needed to extend CPlantBox by introducing a new leaf geometrization that results in realistic leafs. All parameterized geometries of the plant are directly provided by the plant model. In the Unreal Engine, it is possible to alter the environment. WebRTC enables the streaming of the final image composition, which, in turn, can then be directly used to train deep neural networks to increase parameter robustness, for further plant trait detection and validation of original parameters. We enable user-friendly ready-to-use pipelines, providing virtual plant experiment and field visualizations, a python-binding library to access synthetic data and a ready-to-run example to train models.

A framework for modeling, generating, simulating, and predicting carbon dioxide dispersion indoors using cell-DEVS and deep learning

Carbon dioxide concentration in enclosed spaces is an air quality indicator that affects occupants’ well-being. To maintain healthy carbon dioxide levels indoors, enclosed space settings must be adjusted to maximize air quality while minimizing energy consumption. Studying the effect of these settings on carbon dioxide concentration levels is not feasible through physical experimentation and data collection. This problem can be solved by using validated simulation models, generating indoor settings scenarios, simulating those scenarios, and studying results. In previous work, we presented a formal Cellular Discrete Event System Specifications simulation model for studying carbon dioxide dispersion in rooms with various settings. However, designers may need to predict the results of altering large combinations of settings on air quality. Generating and simulating multiple scenarios with different combinations of space settings to test their effect on indoor air quality is time-consuming. In this research, we solve the two problems of the lack of ground truth data and the inefficiency of producing and studying simulation results for many combinations of settings by proposing a novel framework. The framework utilizes a Cellular Discrete Event System Specifications model, simulates different scenarios of enclosed spaces with various settings, and collects simulation results to form a data set to train a deep neural network. Without needing to generate all possible scenarios, the trained deep neural network is used to predict unknown settings of the closed space when other settings are altered. The framework facilitates configuring enclosed spaces to enhance air quality. We illustrate the framework uses through a case study.

A fully automated machine learning-based methodology for personalized radiation dose assessment in thoracic and abdomen CT

PurposeTo develop a machine learning-based methodology for patient-specific radiation dosimetry in thoracic and abdomen CT. MethodsThree hundred and thirty-one thoracoabdominal radiotherapy-planning CT examinations with the respective organ/patient contours were collected retrospectively for the development and validation of segmentation 3D-UNets. Moreover, 97 diagnostic thoracic and 89 diagnostic abdomen CT examinations were collected retrospectively. For each of the diagnostic CT examinations, personalized MC dosimetry was performed. The data derived from MC simulations along with the respective CT data were used for the training and validation of a dose prediction deep neural network (DNN). An algorithm was developed to utilize the trained models and perform patient-specific organ dose estimates for thoracic and abdomen CT examinations. The doses estimated with the DNN were compared with the respective doses derived from MC simulations. A paired t-test was conducted between the DNN and MC results. Furthermore, the time efficiency of the proposed methodology was assessed. ResultsThe mean percentage differences (range) between DNN and MC dose estimates for the lungs, liver, spleen, stomach, and kidneys were 7.2 % (0.2–24.1 %), 5.5 % (0.4–23.0 %), 7.9 % (0.6–22.3 %), 6.9 % (0.0–23.0 %) and 6.7 % (0.3–22.6 %) respectively. The differences between DNN and MC dose estimates were not significant (p-value = 0.12). Moreover, the mean processing time of the proposed workflow was 99 % lower than the respective time needed for MC-based dosimetry. ConclusionsThe proposed methodology can be used for rapid and accurate patient-specific dosimetry in chest and abdomen CT.

Fault Diagnosis for Modular Multilevel Converter (MMC) Based on Deep Learning: An Edge Implementation Using Binary Neural Network

Recently deep neural network (DNN) has been proposed for fault diagnosis of modular multilevel converter (MMC). The training of DNN is convenient due to its end-to-end feature, and DNN has a strong learning ability due to its deep structure. However, DNN is hard to be implemented at the edge (i.e., the MMC side) since it involves a large number of floating-point parameters and calculations. This paper proposes a methodology to deploy a DNN with fault diagnosis purpose at the edge. First, the floating-point DNN is converted to a binary neural network (BNN) version. BNN not only binarizes the floating-point weights as “1” or “–1” to save the storage footprint, but also replaces the floating-point multiply-accumulates (MACs) with bitwise operations to reduce the computation complexity. Second, the computation of different BNN layers is distributed to different embedded real-time controllers in MMC, so as to fully use the existing computing resources. With the proposed methodology, a typical DNN, namely one-dimensional convolutional neural network (1-D CNN) for open-circuit fault (OCF) diagnosis of MMC, which has 238.93kB parameter storage footprint and 91,752 floating-point MACs, is replaced by a BNN with only 7.48kB parameter storage footprint and bitwise operations. 96.87% storage footprint is saved and almost floating-point operations are replaced, and as a tradeoff, the diagnosis accuracy is only reduced by 4.33% compared with the 1-D CNN counterpart. This work shows the huge potential of edge intelligence for fault diagnosis of power electronics.

Scaling the Inference of Digital Pathology Deep Learning Models Using CPU-Based High-Performance Computing

Digital pathology whole-slide images (WSIs) are large-size gigapixel images and image analysis based on deep learning artificial intelligence (AI) technology often involves pixel-wise testing of a trained deep learning neural network (DLNN) on hundreds of WSI images, which is time consuming. We take advantage of High-Performance Computing (HPC) facilities to parallelize this procedure into multiple independent (and hence delightfully parallel) tasks. However, traditional software parallelization techniques and regular file formats can have significant scaling problems on HPC clusters. In this work, a useful computational strategy is designed to localize and extract relevant patches in WSI files and group them in HDF5 files well suited for parallel I/O. HPC's array job facilities are adapted for hierarchical scaling and parallelization of WSI pre-processing and testing of trained algorithms. Applying these techniques to testing a trained DLNN on the CAMELYON datasets with 399 WSIs reduced the theoretical processing time of 18 years on a single CPU or 30 days on a single GPU to less than 45 hours on an HPC cluster of 4,000 CPU cores. The efficiency-accuracy trade-off we demonstrated on this dataset further reinforced the importance of efficient computation techniques, without which the accuracy may be sacrificed. The framework developed here for testing DLNNs does not rely on any specific neural network architecture and HPC cluster setup and can be utilized for any large-scale image processing and big-data analysis.

Anomaly detection and visualization for sound data with pre-trained deep neural networks

Anomaly detection in the industrial domain is a critical issue. Although anomaly detection needs expert's experiences or huge person-hours, recent researches of deep neural networks can save much labor. This paper proposes an anomaly detection method for sound data with pre-trained deep neural networks, and an anomaly visualization method to specify the frequency components of anomalous sounds. Common anomaly detection methods train deep neural networks which can detect anomalies in specific domains. However, training such networks from scratch costs a lot because the neural networks have numerous parameters and use lots of data to improve the performance. To address this problem, anomaly detection using the pre-trained image classification networks has been proposed. This method uses feature representations of classification networks and the Mahalanobis distance as the anomaly scores. In this paper, we propose a method which can apply the above method to sound data and evaluate its performance with a dataset for industrial sound which contains machine sounds under normal and anomalous operating conditions in real factory environments. We also propose a visualization method to specify the frequency components when abnormal sounds occur.

Using the IBM analog in-memory hardware acceleration kit for neural network training and inference

Analog In-Memory Computing (AIMC) is a promising approach to reduce the latency and energy consumption of Deep Neural Network (DNN) inference and training. However, the noisy and non-linear device characteristics and the non-ideal peripheral circuitry in AIMC chips require adapting DNNs to be deployed on such hardware to achieve equivalent accuracy to digital computing. In this Tutorial, we provide a deep dive into how such adaptations can be achieved and evaluated using the recently released IBM Analog Hardware Acceleration Kit (AIHWKit), freely available at https://github.com/IBM/aihwkit. AIHWKit is a Python library that simulates inference and training of DNNs using AIMC. We present an in-depth description of the AIHWKit design, functionality, and best practices to properly perform inference and training. We also present an overview of the Analog AI Cloud Composer, a platform that provides the benefits of using the AIHWKit simulation in a fully managed cloud setting along with physical AIMC hardware access, freely available at https://aihw-composer.draco.res.ibm.com. Finally, we show examples of how users can expand and customize AIHWKit for their own needs. This Tutorial is accompanied by comprehensive Jupyter Notebook code examples that can be run using AIHWKit, which can be downloaded from https://github.com/IBM/aihwkit/tree/master/notebooks/tutorial.

Federated Learning: Centralized and P2P for a Siamese Deep Learning Model for Diabetes Foot Ulcer Classification

It is a known fact that AI models need massive amounts of data for training. In the medical field, the data are not necessarily available at a single site but are distributed over several sites. In the field of medical data sharing, particularly among healthcare institutions, the need to maintain the confidentiality of sensitive information often restricts the comprehensive utilization of real-world data in machine learning. To address this challenge, our study experiments with an innovative approach using federated learning to enable collaborative model training without compromising data confidentiality and privacy. We present an adaptation of the federated averaging algorithm, a predominant centralized learning algorithm, to a peer-to-peer federated learning environment. This adaptation led to the development of two extended algorithms: Federated Averaging Peer-to-Peer and Federated Stochastic Gradient Descent Peer-to-Peer. These algorithms were applied to train deep neural network models for the detection and monitoring of diabetic foot ulcers, a critical health condition among diabetic patients. This study compares the performance of Federated Averaging Peer-to-Peer and Federated Stochastic Gradient Descent Peer-to-Peer with their centralized counterparts in terms of model convergence and communication costs. Additionally, we explore enhancements to these algorithms using targeted heuristics based on client identities and f1-scores for each class. The results indicate that models utilizing peer-to-peer federated averaging achieve a level of convergence that is comparable to that of models trained via conventional centralized federated learning approaches. This represents a notable progression in the field of ensuring the confidentiality and privacy of medical data for training machine learning models.

Deep learning for the prediction of clinical outcomes in internet-delivered CBT for depression and anxiety.

In treating depression and anxiety, just over half of all clients respond. Monitoring and obtaining early client feedback can allow for rapidly adapted treatment delivery and improve outcomes. This study seeks to develop a state-of-the-art deep-learning framework for predicting clinical outcomes in internet-delivered Cognitive Behavioural Therapy (iCBT) by leveraging large-scale, high-dimensional time-series data of client-reported mental health symptoms and platform interaction data. We use de-identified data from 45,876 clients on SilverCloud Health, a digital platform for the psychological treatment of depression and anxiety. We train deep recurrent neural network (RNN) models to predict whether a client will show reliable improvement by the end of treatment using clinical measures, interaction data with the iCBT program, or both. Outcomes are based on total improvement in symptoms of depression (Patient Health Questionnaire-9, PHQ-9) and anxiety (Generalized Anxiety Disorder-7, GAD-7), as reported within the iCBT program. Using internal and external datasets, we compare the proposed models against several benchmarks and rigorously evaluate them according to their predictive accuracy, sensitivity, specificity and AUROC over treatment. Our proposed RNN models consistently predict reliable improvement in PHQ-9 and GAD-7, using past clinical measures alone, with above 87% accuracy and 0.89 AUROC after three or more review periods, outperforming all benchmark models. Additional evaluations demonstrate the robustness of the achieved models across (i) different health services; (ii) geographic locations; (iii) iCBT programs, and (iv) client severity subgroups. Results demonstrate the robust performance of dynamic prediction models that can yield clinically helpful prognostic information ready for implementation within iCBT systems to support timely decision-making and treatment adjustments by iCBT clinical supporters towards improved client outcomes.

Backpropagation-free training of deep physical neural networks.

Recent successes in deep learning for vision and natural language processing are attributed to larger models but come with energy consumption and scalability issues. Current training of digital deep learning models primarily relies on backpropagation that is unsuitable for physical implementation. Here, we proposed a simple deep neural network architecture augmented by a physical local learning (PhyLL) algorithm, enabling supervised and unsupervised training of deep physical neural networks, without detailed knowledge of the nonlinear physical layer's properties. We trained diverse wave-based physical neural networks in vowel and image classification experiments, showcasing our approach's universality. Our method shows advantages over other hardware-aware training schemes by improving training speed, enhancing robustness, and reducing power consumption by eliminating the need for system modelling and thus decreasing digital computation.

Using deep neural networks as a guide for modeling human planning

When developing models in cognitive science, researchers typically start with their own intuitions about human behavior in a given task and then build in mechanisms that explain additional aspects of the data. This refinement step is often hindered by how difficult it is to distinguish the unpredictable randomness of people’s decisions from meaningful deviations between those decisions and the model. One solution for this problem is to compare the model against deep neural networks trained on behavioral data, which can detect almost any pattern given sufficient data. Here, we apply this method to the domain of planning with a heuristic search model for human play in 4-in-a-row, a combinatorial game where participants think multiple steps into the future. Using a data set consisting of 10,874,547 games, we train deep neural networks to predict human moves and find that they accurately do so while capturing meaningful patterns in the data. Thus, deviations between the model and the best network allow us to identify opportunities for model improvement despite starting with a model that has undergone substantial testing in previous work. Based on this analysis, we add three extensions to the model that range from a simple opening bias to specific adjustments regarding endgame planning. Overall, our work demonstrates the advantages of model comparison with a high-performance deep neural network as well as the feasibility of scaling cognitive models to massive data sets for systematically investigating the processes underlying human sequential decision-making.

Relevant input data for crack feature segmentation with deep learning on SEM imagery and topography data

Fractography plays a critical role in failure analysis of engineering components and has a considerable importance for safety investigations. Usually, the interpretation of fracture surfaces is done by experts with the help of literature and experimental data, that requires a lot of experience. The use of deep learning (DL) with neural networks in failure analysis becomes more and more relevant with the rapidly developing possibilities. Especially, the modern network architectures can assist fractographers in determining various fracture features on SEM images of the fracture surfaces. The basis for the best possible evaluation is the understanding of the influence of the input data used for training deep neural networks (DNN). Therefore, this study discusses the influence of the selection of the input data used for the prediction quality of these networks in order to take this into account for future data acquisition. Specimens of various metallic materials were subjected to fatigue cracking experiment under laboratory conditions. The fractured surfaces were then imaged using various modes or detectors (such as SE, BSE and topography) in SEM, and those captured images were used to create a training data set. The relevance of the individual data for the quality of the prediction is determined by a specific combination of the different detector data. For the training, the well-established architecture of a UNet-ResNet34 with a fixed set of hyperparameters is used. It has been found in this present study that the combination of all input data significantly increases the prediction accuracy, whereby even the combination of SE and BSE data provides considerable advantages over the exclusive use of SE images.

Multi-Adaptive Optimization for multi-task learning with deep neural networks

Multi-task learning is a promising paradigm to leverage task interrelations during the training of deep neural networks. A key challenge in the training of multi-task networks is to adequately balance the complementary supervisory signals of multiple tasks. In that regard, although several task-balancing approaches have been proposed, they are usually limited by the use of per-task weighting schemes and do not completely address the uneven contribution of the different tasks to the network training. In contrast to classical approaches, we propose a novel Multi-Adaptive Optimization (MAO) strategy that dynamically adjusts the contribution of each task to the training of each individual parameter in the network. This automatically produces a balanced learning across tasks and across parameters, throughout the whole training and for any number of tasks. To validate our proposal, we perform comparative experiments on real-world datasets for computer vision, considering different experimental settings. These experiments allow us to analyze the performance obtained in several multi-task scenarios along with the learning balance across tasks, network layers and training steps. The results demonstrate that MAO outperforms previous task-balancing alternatives. Additionally, the performed analyses provide insights that allow us to comprehend the advantages of this novel approach for multi-task learning.

Theory-inspired machine learning for stress–strain curve prediction of short fiber-reinforced composites with unseen design space

Computational analysis of composites is well established; however, the process is time-consuming, owing to the disparate length scales and nonlinearities involved. Therefore, the use of machine learning models to predict the mechanical properties of short fiber-reinforced composites (SFRCs) is promising when large amounts of data are lacking. However, predicting data for an unseen design space remains challenging. The use of existing theoretical models for training is a viable solution to address this issue. This paper proposes a theory-guided machine learning (TGML) framework through training deep neural networks, based on the finite element analysis dataset that considers the cohesive effect of the interface, to predict the nonlinear mechanical responses of SFRCs. Our results demonstrate that incorporating the Halpin-Tsai theory with machine learning improves the predictive performance of the model in an unseen design space. Moreover, inspired by this theory, we propose a theory-inspired two-phase machine learning (TPML) approach to further improve predictive performance. Our results indicate that TGML and TPML capture more information from the data and thus enhance predictive performance. The proposed method can be adapted to the analyses of other composites with nonlinear mechanical behaviors.

Hierarchical waste detection with weakly supervised segmentation in images from recycling plants

Reducing environmental pollution with household waste and emissions from the computing clusters is an urgent technological problem. In our work, we explore both of these aspects: the deep learning application to improve the efficiency of waste recognition on recycling plant’s conveyor, as well as carbon dioxide emission from the computing devices used in this process. To conduct research, we developed an unique open WaRP dataset that demonstrates the best diversity among similar industrial datasets and contains more than 10,000 images with 28 different types of recyclable goods (bottles, glasses, card boards, cans, detergents, and canisters). Objects can overlap, be in poor lighting conditions, or significantly distorted. On the WaRP dataset, we study training and evaluation of cutting-edge deep neural networks for detection, classification and segmentation tasks. Additionally, we developed a hierarchical neural network approach called H-YC with weakly supervised waste segmentation. It provided a notable increase in the detection quality and made it possible to segment images, learning only having class labels, not their masks. Both the suggested hierarchical approach and the WaRP dataset have shown great industrial application potential.

Transfer learning for deep neural networks-based classification of breast cancer X-ray images

ABSTRACT Nowadays, deep neural networks (DNNs) are helpful tools for mammogram classification in breast cancer screening. But, in Vietnam, there is a relatively small number of mammograms for training DNNs. Therefore, this study aims to apply transfer learning techniques to improve the performance of DNN models. In the first step, 10,418 breast cancer images from the Digital Database for Screening Mammography were used for training the CNN model ResNet 34. In the second step, we fine-tune this model on the Hanoi Medical University (HMU) database with 6,248 Vietnamese mammograms. The optimal model of ResNet 34 among these models achieves a macAUC of 0.766 in classifying breast cancer X-ray images into three Breast Imaging-Reporting and Data System (BI-RADS) categories, BI-RADS 045 (‘incomplete and malignance’), BI-RADS 1 (‘normal’), and BI-RADS 23 (‘benign’), when tested on the test dataset. This result is higher than the result of the ResNet 50 model trained only on the X-ray dataset of 7,912 breast cancer images of the HMU dataset, which achieves a macAUC of 0.754. A comparison of the performance of the proposed model of ResNet 34 applying transfer learning with other works shows that our model’s evaluation results are higher than those of the compared models.

Machine learning-based optimization of segmented thermoelectric power generators using temperature-dependent performance properties

Segmented thermoelectric generators (STEGs) provide an excellent platform for thermal energy harvesting devices because they improve power generation performance across a broad range of operating temperatures. Despite the benefit of direct thermal energy-to-electricity conversion, conventional STEG optimization approaches are unable to provide a systematic method for selecting the optimal multiple stacks of p- and-n-type thermoelectric materials (TEs) legs from a set of numerous TE materials. In this study, we propose a systematic optimization method based on machine learning to find the optimal STEG for performance maximization. A deep neural network (DNN) is trained using an initial dataset generated via the Finite Element Method (FEM), with inputs including the temperature-dependent properties of p- and n-type TE materials, the lengths of each segment, and external loads, as well as the corresponding performance outputs. The trained DNN captures the inherent nonlinear relationship between these inputs and outputs. The combination of a trained DNN and a genetic algorithm (GA) efficiently navigates a vast design space using 88 p-type and 70 n-type TE materials along with device factors. It formulates four stacked segment pairs in p- and n-leg TEGs, targeting new superior designs with enhanced power, efficiency, or both. The DNN is iteratively refined via active learning (AL) by incorporating new superior STEG designs to enhance prediction accuracy. The optimized STEGs exhibit power and efficiency that are 1.91 and 1.5 times higher, respectively, than the top designs from the initial training dataset composed of 157,916 STEGs. Furthermore, compared to conventional TEG designs without segmentation, our method discovered numerous high-performing STEG designs.

Infrared UAV Target Detection Based on Continuous-Coupled Neural Network.

The task of the detection of unmanned aerial vehicles (UAVs) is of great significance to social communication security. Infrared detection technology has the advantage of not being interfered with by environmental and other factors and can detect UAVs in complex environments. Since infrared detection equipment is expensive and data collection is difficult, there are few existing UAV-based infrared images, making it difficult to train deep neural networks; in addition, there are background clutter and noise in infrared images, such as heavy clouds, buildings, etc. The signal-to-clutter ratio is low, and the signal-to-noise ratio is low. Therefore, it is difficult to achieve the UAV detection task using traditional methods. The above challenges make infrared UAV detection a difficult task. In order to solve the above problems, this work drew upon the visual processing mechanism of the human brain to propose an effective framework for UAV detection in infrared images. The framework first determines the relevant parameters of the continuous-coupled neural network (CCNN) through the image's standard deviation, mean, etc. Then, it inputs the image into the CCNN, groups the pixels through iteration, then obtains the segmentation result through expansion and erosion, and finally, obtains the final result through the minimum circumscribed rectangle. The experimental results showed that, compared with the existing most-advanced brain-inspired image-understanding methods, this framework has the best intersection over union (IoU) (the intersection over union is the overlapping area between the predicted segmentation and the label divided by the joint area between the predicted segmentation and the label) in UAV infrared images, with an average of 74.79% (up to 97.01%), and can effectively realize the task of UAV detection.

Just-in-time deep learning for real-time X-ray computed tomography

Real-time X-ray tomography pipelines, such as implemented by RECAST3D, compute and visualize tomographic reconstructions in milliseconds, and enable the observation of dynamic experiments in synchrotron beamlines and laboratory scanners. For extending real-time reconstruction by image processing and analysis components, Deep Neural Networks (DNNs) are a promising technology, due to their strong performance and much faster run-times compared to conventional algorithms. DNNs may prevent experiment repetition by simplifying real-time steering and optimization of the ongoing experiment. The main challenge of integrating DNNs into real-time tomography pipelines, however, is that they need to learn their task from representative data before the start of the experiment. In scientific environments, such training data may not exist, and other uncertain and variable factors, such as the set-up configuration, reconstruction parameters, or user interaction, cannot easily be anticipated beforehand, either. To overcome these problems, we developed just-in-time learning, an online DNN training strategy that takes advantage of the spatio-temporal continuity of consecutive reconstructions in the tomographic pipeline. This allows training and deploying comparatively small DNNs during the experiment. We provide software implementations, and study the feasibility and challenges of the approach by training the self-supervised Noise2Inverse denoising task with X-ray data replayed from real-world dynamic experiments.