Size Of Training Data Research Articles

Long-short term memory networks aided fault detection of power facilities

In recent years, long-short term memory (LSTM) networks have emerged as a promising tool for time-series analysis, offering the ability to capture temporal dependencies effectively, especially for the fault detection in power facilities which is crucial for maintaining operational integrity and ensuring continuous power supply. This paper presents a study on the utilization of LSTM networks for fault detection in power facilities, focusing on evaluating the detection accuracy as a performance metric. We propose a novel LSTM-based fault detection framework that utilizes historical operational data to identify abnormal patterns indicative of faults or anomalies. The framework is trained on labeled data to learn the normal operational behavior of the power facility, enabling it to detect deviations from expected patterns in practical data streams. To assess the effectiveness of the proposed approach, extensive experiments are conducted using real-world operational data from diverse power facilities. We systematically evaluate the detection accuracy of the LSTM-based fault detection system under various operating conditions, including different types of faults and levels of noise in the data. Comparative analyses are performed against traditional fault detection methods to demonstrate the superiority of the LSTM-based approach in terms of accuracy. Moreover, sensitivity analyses are conducted to investigate the impact of key parameters, such as network architecture and training data size, on the detection performance. The results illustrate the robustness and effectiveness of the proposed LSTM-based fault detection framework, highlighting its potential for enhancing the reliability and efficiency of power facility operations.

Bayesian dynamic modelling for probabilistic prediction of pavement condition

Significant funds have been allocated to maintain road networks each year in developed countries. Performance prediction is crucial for pavement management systems to adjust working plans and budget allocation. As a Bayesian nonparametric method, Gaussian process regression (GPR) is powerful in predicting nonlinear time series and quantifying uncertainty. However, it remains computationally intensive and fails to adapt to the time-varying characteristics. To address such issues, a dynamic GPR model is proposed for probabilistic prediction of the International Roughness Index (IRI) for flexible pavements. A moving window strategy is developed to substantially shrink the size of training data, which effectively alleviates computational cost and thus leads to a dynamic GPR. A genetic algorithm is then adopted to determine the optimal window size by considering the trade-off between computational efficiency and accuracy. A dataset acquired from Long-Term Pavement Performance (LTPP) is used to demonstrate the feasibility of the dynamic GPR. Its performance is compared to traditional GPR as well as dynamic and static Bayesian linear regression (BLR) models. The comparison results indicate that the proposed dynamic GPR can increase the accuracy by 0.86, 1.52, and 2.27 times for dynamic BLR, static GPR, and static BLR, respectively. It exhibits the best results in terms of accuracy and uncertainty metrics due to its nonlinear modelling and time-varying ability.

Transfer learning enables identification of multiple types of RNA modifications using nanopore direct RNA sequencing

Nanopore direct RNA sequencing (DRS) has emerged as a powerful tool for RNA modification identification. However, concurrently detecting multiple types of modifications in a single DRS sample remains a challenge. Here, we develop TandemMod, a transferable deep learning framework capable of detecting multiple types of RNA modifications in single DRS data. To train high-performance TandemMod models, we generate in vitro epitranscriptome datasets from cDNA libraries, containing thousands of transcripts labeled with various types of RNA modifications. We validate the performance of TandemMod on both in vitro transcripts and in vivo human cell lines, confirming its high accuracy for profiling m6A and m5C modification sites. Furthermore, we perform transfer learning for identifying other modifications such as m7G, Ψ, and inosine, significantly reducing training data size and running time without compromising performance. Finally, we apply TandemMod to identify 3 types of RNA modifications in rice grown in different environments, demonstrating its applicability across species and conditions. In summary, we provide a resource with ground-truth labels that can serve as benchmark datasets for nanopore-based modification identification methods, and TandemMod for identifying diverse RNA modifications using a single DRS sample.

Deriving Input Variables through Applied Machine Learning for Short-Term Electric Load Forecasting in Eskilstuna, Sweden

As the demand for electricity, electrification, and renewable energy rises, accurate forecasting and flexible energy management become imperative. Distribution network operators face capacity limits set by regional grids, risking economic penalties if exceeded. This study examined data-driven approaches of load forecasting to address these challenges on a city scale through a use case study of Eskilstuna, Sweden. Multiple Linear Regression was used to model electric load data, identifying key calendar and meteorological variables through a rolling origin validation process, using three years of historical data. Despite its low cost, Multiple Linear Regression outperforms the more expensive non-linear Light Gradient Boosting Machine, and both outperform the “weekly Naïve” benchmark with a relative Root Mean Square Errors of 32–34% and 39–40%, respectively. Best-practice hyperparameter settings were derived, and they emphasize frequent re-training, maximizing the training data size, and setting a lag size larger than or equal to the forecast horizon for improved accuracy. Combining both models into an ensemble could the enhance accuracy. This paper demonstrates that robust load forecasts can be achieved by leveraging domain knowledge and statistical analysis, utilizing readily available machine learning libraries. The methodology for achieving this is presented within the paper. These models have the potential for economic optimization and load-shifting strategies, offering valuable insights into sustainable energy management.

Transition role of entangled data in quantum machine learning

Entanglement serves as the resource to empower quantum computing. Recent progress has highlighted its positive impact on learning quantum dynamics, wherein the integration of entanglement into quantum operations or measurements of quantum machine learning (QML) models leads to substantial reductions in training data size, surpassing a specified prediction error threshold. However, an analytical understanding of how the entanglement degree in data affects model performance remains elusive. In this study, we address this knowledge gap by establishing a quantum no-free-lunch (NFL) theorem for learning quantum dynamics using entangled data. Contrary to previous findings, we prove that the impact of entangled data on prediction error exhibits a dual effect, depending on the number of permitted measurements. With a sufficient number of measurements, increasing the entanglement of training data consistently reduces the prediction error or decreases the required size of the training data to achieve the same prediction error. Conversely, when few measurements are allowed, employing highly entangled data could lead to an increased prediction error. The achieved results provide critical guidance for designing advanced QML protocols, especially for those tailored for execution on early-stage quantum computers with limited access to quantum resources.

COSST: Multi-Organ Segmentation With Partially Labeled Datasets Using Comprehensive Supervisions and Self-Training.

Deep learning models have demonstrated remarkable success in multi-organ segmentation but typically require large-scale datasets with all organs of interest annotated. However, medical image datasets are often low in sample size and only partially labeled, i.e., only a subset of organs are annotated. Therefore, it is crucial to investigate how to learn a unified model on the available partially labeled datasets to leverage their synergistic potential. In this paper, we systematically investigate the partial-label segmentation problem with theoretical and empirical analyses on the prior techniques. We revisit the problem from a perspective of partial label supervision signals and identify two signals derived from ground truth and one from pseudo labels. We propose a novel two-stage framework termed COSST, which effectively and efficiently integrates comprehensive supervision signals with self-training. Concretely, we first train an initial unified model using two ground truth-based signals and then iteratively incorporate the pseudo label signal to the initial model using self-training. To mitigate performance degradation caused by unreliable pseudo labels, we assess the reliability of pseudo labels via outlier detection in latent space and exclude the most unreliable pseudo labels from each self-training iteration. Extensive experiments are conducted on one public and three private partial-label segmentation tasks over 12 CT datasets. Experimental results show that our proposed COSST achieves significant improvement over the baseline method, i.e., individual networks trained on each partially labeled dataset. Compared to the state-of-the-art partial-label segmentation methods, COSST demonstrates consistent superior performance on various segmentation tasks and with different training data sizes.

Utilizing convolutional neural networks (CNN) and U-Net architecture for precise crop and weed segmentation in agricultural imagery: A deep learning approach

This study presents the implementation and evaluation of a convolutional neural network (CNN) based image segmentation model using the U-Net architecture for forest image segmentation. The proposed algorithm starts by preprocessing the datasets of satellite images and corresponding masks from a repository source. Data preprocessing involves resizing, normalizing, and splitting the images and masks into training and testing datasets. The U-Net model architecture, comprising encoder and decoder parts with skip connections, is defined and compiled with binary cross-entropy loss and Adam optimizer. Training includes early stopping and checkpoint saving mechanisms to prevent overfitting and retain the best model weights. Evaluation metrics such as Intersection over Union (IoU), Dice coefficient, pixel accuracy, precision, recall, specificity, and F1-score are computed to assess the model's performance. Visualization of results includes comparing predicted segmentation masks with ground truth masks for qualitative analysis. The study emphasizes the importance of training data size in achieving accurate segmentation models and highlights the potential of U-Net architecture for forest image segmentation tasks.

Exploring the Performance of Linear and Nonlinear Models of Time-of-Flight Secondary Ion Mass Spectrometry Spectra.

Multivariate statistical tools and machine learning (ML) techniques can deconvolute hyperspectral data and control the disparity between the number of samples and features in materials science. Nevertheless, the importance of generating sufficient high-quality sample replicates in training data cannot be overlooked, as it fundamentally affects the performance of ML models. Here, we present a quantitative analysis of time-of-flight secondary ion mass spectrometry (ToF-SIMS) spectra of a simple microarray system of two food dyes using partial least-squares (PLS, linear) and random forest (RF, nonlinear) algorithms. This microarray was generated by a high-throughput sample preparation and analysis workflow for fast and efficient acquisition of quality and reproducible spectra via ToF-SIMS. We drew insights from the bias-variance trade-off, investigated the performances of PLS and RF regression models as a function of training data size, and inferred the amount of data needed to construct accurate and reliable regression models. In addition, we found that the spectral concatenation of positive and negative ToF-SIMS spectra improved the model performances. This study provides an empirical basis for future design of high-throughput microarrays and multicomponent systems, for the purpose of analysis with ToF-SIMS and ML.

Predicting Cyber Security Threats in The Digital Age: A Systematic Review and Analysis

Insider threats pose a significant challenge in securing today's digital economy, particularly within the convergence of Critical Infrastructure (CI) and Industry 4.0 applications. This study examines current trends, challenges, and algorithms for detecting and predicting insider threats. We analyse key issues such as data non-stationarity and collusion attacks. The paper proposes a model to address these challenges and evaluates its effectiveness. While the initial results show a 66% accuracy in terms of False Positive Rate (FPR), the study recommends doubling the training data size to further enhance the model's prediction accuracy. This research contributes to improved security solutions for safeguarding critical infrastructure in the era of Industry 4.0. In today's digital world, even simple mistakes when using online systems can lead to cyberattacks with serious consequences. These consequences can include damage to an organization's reputation, financial losses, and disruptions to daily operations. This problem is especially concerning in developing countries. Researchers tackled this issue by building a machine learning model that predicts who might fall victim to cyberattacks. The model analyses social and economic data to identify key factors that make someone more susceptible. This model achieved a high level of accuracy by utilizing a sophisticated machine learning technique combined with an algorithm that discovers hidden patterns in data. To train the model, researchers collected information from both victims and those who haven't been attacked. They then employed a special technique to expand the data available for training, ultimately generating valuable insights that can help predict cyberattacks and the associated risks.

Prediction improvement of compressive strength and strain of directionally solidified TiAl alloy based on training data size adjustment

Compressive strength and compressive strain, which are important mechanical properties of directionally solidified TiAl alloy, were predicted using machine learning algorithms, specifically Multiple Linear Regression (MLR), and Random Forest Regression (RFR). The input variables for the machine learning model were designated as the composition of the directionally solidified TiAl alloy, experimental parameters (pulling velocity, input power), and compression test parameters (strain rate, compression temperature). Compressive strength and compressive strain were designated as the output variables. Although the typical ratio of training and test data is 8:2, this study used different ratios of 9:1, 7:3, and 6:4 for machine learning, and excellent R2 values were obtained for all ratios. The feature importance, which can identify the factor that has the most influence on the output variables, was obtained through the RFR algorithm. According to the feature importance, temperature was found to have the greatest influence on compressive strength, while the Erbium (Er) element had the most significant influence on compressive strain. Through the results of feature importance, it was possible to quantitatively investigate the relationship between the Er element, a microalloying element that affects the microstructure of TiAl alloy, and the compressive properties. Furthermore, the study was conducted on which data ratio between training and test data is most suitable for predicting the compressive strength and strain of TiAl alloy.

Enabling action crossmodality for a pretrained large language model

Natural language processing and vision tasks have recently seen large improvements through the rise of Transformer architectures. The high-performing large language models (LLMs) benefit from large textual datasets that are numerously available online. However, action and bidirectional action-language tasks are less developed, as these require more specific and labeled data. Therefore, we aim at enabling these robotic action capabilities for a pretrained LLM, while maintaining high efficiency with regards to the required training time and data size. To achieve this, we split up a Transformer-based LLM and insert a multimodal architecture into it. Specifically, we split a pretrained T5 LLM between its encoder and decoder parts, to insert a crossmodal Transformer component of a Paired Transformed Autoencoders (PTAE) bidirectional action-language model. The experiments are conducted on a new dataset, consisting of unimodal language translation and crossmodal bidirectional action-language translation. The natural language capabilities of the original T5 are re-established efficiently by training the crossmodal Transformer, which requires only one 5.7 millionth of the T5 model’s original training data. Furthermore, the new model, called CrossT5, achieves high accuracy for the vision- and language-guided robotic action tasks. By design, the CrossT5 agent acts robustly when tested with language commands not included in the dataset. The results demonstrate that this novel approach is successful in combining the advanced linguistic capabilities of LLMs with the low-level robotic control skills of vision-action models. The code is available at this URL: https://github.com/samsoneko/CrossT5.

Overcoming the semantic gap in the customer-to-manufacturer (C2M) platform: A soft prompts-based approach with pretrained language models

Customer-to-Manufacturer (C2M) is a strategy in smart manufacturing where customers collaborate with manufacturers for customized product development on an online platform. The platform enables the shift from the traditional manufacturing process, which is driven by research and marketing, toward a customer-centric product development process. However, a challenge arises as customers lack technical knowledge to communicate their product specifications effectively, creating a semantic gap. This paper proposes a soft prompt-based network structure that utilizes pretrained language models to bridge the semantic gap on the C2M platform. To address limited customer needs data and imbalanced classes, a large corpus of product review texts is used to establish a mapping between reviews and product specifications. A smaller set of customer needs text is then employed to adapt this mapping to the target customer needs-product specifications relationship, thereby closing the semantic gap. The experimental results demonstrate the effectiveness of the proposed model adaptation operation and the prompting structure. Additionally, the experiments highlight the robustness of the proposed method against variations in training data size, thereby mitigating the challenges posed by imbalanced classes. The proposed method could potentially bring innovation to product customization and C2M platform development. By bridging the semantic gap, companies can better integrate customers in the co-design process and effectively translate customer needs into actionable product specifications.

Bayesian approach to assessing population differences in genetic risk of disease with application to prostate cancer.

Population differences in risk of disease are common, but the potential genetic basis for these differences is not well understood. A standard approach is to compare genetic risk across populations by testing for mean differences in polygenic scores, but existing studies that use this approach do not account for statistical noise in effect estimates (i.e., the GWAS betas) that arise due to the finite sample size of GWAS training data. Here, we show using Bayesian polygenic score methods that the level of uncertainty in estimates of genetic risk differences across populations is highly dependent on the GWAS training sample size, the polygenicity (number of causal variants), and genetic distance (FST) between the populations considered. We derive a Wald test for formally assessing the difference in genetic risk across populations, which we show to have calibrated type 1 error rates under a simplified assumption that all SNPs are independent, which we achieve in practise using linkage disequilibrium (LD) pruning. We further provide closed-form expressions for assessing the uncertainty in estimates of relative genetic risk across populations under the special case of an infinitesimal genetic architecture. We suggest that for many complex traits and diseases, particularly those with more polygenic architectures, current GWAS sample sizes are insufficient to detect moderate differences in genetic risk across populations, though more substantial differences in relative genetic risk (relative risk > 1.5) can be detected. We show that conventional approaches that do not account for sampling error from the training sample, such as using a simple t-test, have very high type 1 error rates. When applying our approach to prostate cancer, we demonstrate a higher genetic risk in African Ancestry men, with lower risk in men of European followed by East Asian ancestry.

Deep learning reconstruction of high-Reynolds-number turbulent flow field around a cylinder based on limited sensors

It is of significant importance to reconstruct the high-dimensional global flow field using limited sensor data, considering that sensors are often local and limited in full-scale measurements. In the literature, the reconstruction of laminar flows at low Reynolds number (Re) was mostly investigated. This study focuses on high-Re turbulent flows around a body, and aims to systematically investigate the reconstruction performance of deep learning models, including multilayer perceptron (MLP), convolutional neural network (CNN), and generative adversarial neural network (GAN). The extensive assessment encompasses the prediction accuracy of the instantaneous, phase-averaged, and mean flow fields, and the training efficiency. Results shows that GAN demonstrates a obvious advantage in reconstructing small-scale vortex structures. MLP is slightly better at reconstructing time-averaged flow fields (3%) and has a significant advantage in reconstruction efficiency. We also examine the generalization ability by applying transfer learning at different Re. Additionally, sensitivity of sensor number, training data size, and optimization of sensor layout are studied to further augment the accuracy of flow field reconstruction.

Building Shortcuts between Distant Nodes with Biaffine Mapping for Graph Convolutional Networks

Multiple recent studies show a paradox in graph convolutional networks (GCNs)—that is, shallow architectures limit the capability of learning information from high-order neighbors, whereas deep architectures suffer from over-smoothing or over-squashing. To enjoy the simplicity of shallow architectures and overcome their limits of neighborhood extension, in this work we introduce a biaffine technique to improve the expressiveness of GCNs with a shallow architecture. The core design of our method is to learn direct dependency on long-distance neighbors for nodes, with which only 1-hop message passing is capable of capturing rich information for node representation. Besides, we propose a multi-view contrastive learning method to exploit the representations learned from long-distance dependencies. Extensive experiments on nine graph benchmark datasets suggest that the shallow biaffine graph convolutional networks (BAGCN) significantly outperform state-of-the-art GCNs (with deep or shallow architectures) on semi-supervised node classification. We further verify the effectiveness of biaffine design in node representation learning and the performance consistency on different sizes of training data.

Adaptive sampling physics-informed neural network method for high-order rogue waves and parameters discovery of the (2 + 1)-dimensional CHKP equation.

Rogue waves are important physical phenomena, which have wide applications in nonlinear optics, hydrodynamics, Bose-Einstein condensates, and oceanic and atmospheric dynamics. We find that when using the original PINNs to study rogue waves of high dimensional PDEs, the prediction performance will become very poor, especially for high-order rogue waves due to that the randomness of selection of sample points makes insufficient use of the physical information describing the local sharp regions of rogue waves. In this paper, we propose an adaptive sampling physics-informed neural network method (ASPINN), which renders the points in local sharp regions to be selected sufficiently by a new adaptive search algorithm to lead to a prefect prediction performance. To valid the performance of our method, the (2+1)-dimensional CHKP equation is taken as an illustrative example. Experimental results reveal that the original PINNs can hardly be able to predict dynamical behaviors of the high-order rogue waves for the CHKP equation, but the ASPINN method can not only predict dynamical behaviors of these high-order rogue waves, but also greatly improve the prediction efficiency and accuracy to four orders of magnitude. Then, the data-driven inverse problem for the CHKP equation with different levels of corrupted noise is studied to show that the ASPINN method has good robustness. Moreover, some main factors affecting the neural network performance are discussed in detail, including the size of training data, the number of layers of the neural network, and the number of neurons per layer.

AI-Based Design With Data Trimming for Hybrid Phase Shift Modulation for Minimum-Current-Stress Dual Active Bridge Converter

Dual active bridge (DAB) topology has been recognized as the key circuit for the next generation of high-frequency-link power conversion systems. In order to realize desired operating performance, DAB modulation strategies need to be selected carefully. Different modulation strategies have been considered and combined into a hybrid one, which is able to fully optimize performance. However, to develop a hybrid modulation strategy, the conventional methods including harmonic model and piecewise model have difficulty in balancing modeling accuracy and manpower burden. Although recent data-driven modulation approaches can automate the design process, the accuracy of data-driven models decreases nontrivially with the existence of outliers and without sufficient data. Hence, to alleviate human-dependence and improve modeling accuracy, this paper proposes an AI-based design with data trimming (AI-DT) for hybrid phase shift modulation. Two two-degree-of-freedom modulation strategies are considered for the sake of optimal current stress performance. AI-DT firstly adopts the one-class support vector machine (SVM) to exclude outliers. Second, the state-of-the-art extreme gradient boosting (XGBoost), which is insensitive to training data size, is adopted to build data-driven current stress models for the DAB converter. After that, differential evolution (DE) algorithm helps to choose modulation strategy and optimize modulation parameters for optimal current stress. Generally, the proposed AI-DT is developed in an automated fashion, which largely relieves manual computational complexity while exhibiting satisfactory modulation accuracy. The effectiveness of the proposed AI-DT approach has been experimentally verified with a 1kW hardware prototype, realizing optimal current stress under full operating conditions with more than 96.5% as peak efficiency.

Quantifying Seagrass Density Using Sentinel-2 Data and Machine Learning

Seagrasses, rooted aquatic plants growing completely underwater, are extremely important for the coastal ecosystem. They are an important component of the total carbon burial in the ocean, they provide food, shelter, and nursery to many aquatic organisms in coastal ecosystems, and they improve water quality. Due to human activity, seagrass coverage has been rapidly declining, and there is an urgent need to monitor seagrasses consistently. Seagrass coverage has been closely monitored in the Chesapeake Bay since 1970 using air photos and ground samples. These efforts are costly and time-consuming. Many studies have used remote sensing data to identify seagrass bed outlines, but few have mapped seagrass bed density. This study used Sentinel-2 satellite data and machine learning in Google Earth Engine and the Chesapeake Bay Program field data to map seagrass density. We used seagrass density data from the Chincoteague and Sinepuxent Bay to train machine learning algorithms and evaluate their accuracies. Out of the four machine learning models tested (Naive Bayes (NB), Classification and Regression Trees (CART), Support Vector Machine (SVM), and Random Forest (RF)), the RF model outperformed the other three models with overall accuracies of 0.874 and Kappa coefficients of 0.777. The SVM and CART models performed similarly and NB performed the poorest. We tested two different approaches to assess the models’ accuracy. When we used all the available ground samples to train the models, whereby our analysis showed that model performance was associated with seagrass density class, and that higher seagrass density classes had better consumer accuracy, producer accuracy, and F1 scores. However, the association of model performance with seagrass density class disappeared when using the same training data size for each class. Very sparse and dense seagrass classes had replacedhigherbetter accuracies than the sparse and moderate seagrass density classes. This finding suggests that training data impacts machine learning model performance. The uneven training data size for different classes can result in biased assessment results. Selecting proper training data and machine learning models are equally important when using machine learning and remote sensing data to map seagrass density. In summary, this study demonstrates the potential to map seagrass density using satellite data.

Utilization of Synthetic Near-Infrared Spectra via Generative Adversarial Network to Improve Wood Stiffness Prediction.

Near-infrared (NIR) spectroscopy is widely used as a nondestructive evaluation (NDE) tool for predicting wood properties. When deploying NIR models, one faces challenges in ensuring representative training data, which large datasets can mitigate but often at a significant cost. Machine learning and deep learning NIR models are at an even greater disadvantage because they typically require higher sample sizes for training. In this study, NIR spectra were collected to predict the modulus of elasticity (MOE) of southern pine lumber (training set = 573 samples, testing set = 145 samples). To account for the limited size of the training data, this study employed a generative adversarial network (GAN) to generate synthetic NIR spectra. The training dataset was fed into a GAN to generate 313, 573, and 1000 synthetic spectra. The original and enhanced datasets were used to train artificial neural networks (ANNs), convolutional neural networks (CNNs), and light gradient boosting machines (LGBMs) for MOE prediction. Overall, results showed that data augmentation using GAN improved the coefficient of determination (R2) by up to 7.02% and reduced the error of predictions by up to 4.29%. ANNs and CNNs benefited more from synthetic spectra than LGBMs, which only yielded slight improvement. All models showed optimal performance when 313 synthetic spectra were added to the original training data; further additions did not improve model performance because the quality of the datapoints generated by GAN beyond a certain threshold is poor, and one of the main reasons for this can be the size of the initial training data fed into the GAN. LGBMs showed superior performances than ANNs and CNNs on both the original and enhanced training datasets, which highlights the significance of selecting an appropriate machine learning or deep learning model for NIR spectral-data analysis. The results highlighted the positive impact of GAN on the predictive performance of models utilizing NIR spectroscopy as an NDE technique and monitoring tool for wood mechanical-property evaluation. Further studies should investigate the impact of the initial size of training data, the optimal number of generated synthetic spectra, and machine learning or deep learning models that could benefit more from data augmentation using GANs.

Flattening the curve-How to get better results with small deep-mutational-scanning datasets.

Proteins are used in various biotechnological applications, often requiring the optimization of protein properties by introducing specific amino-acid exchanges. Deep mutational scanning (DMS) is an effective high-throughput method for evaluating the effects of these exchanges on protein function. DMS data can then inform the training of a neural network to predict the impact of mutations. Most approaches use some representation of the protein sequence for training and prediction. As proteins are characterized by complex structures and intricate residue interaction networks, directly providing structural information as input reduces the need to learn these features from the data. We introduce a method for encoding protein structures as stacked 2D contact maps, which capture residue interactions, their evolutionary conservation, and mutation-induced interaction changes. Furthermore, we explored techniques to augment neural network training performance on smaller DMS datasets. To validate our approach, we trained three neural network architectures originally used for image analysis on three DMS datasets, and we compared their performances with networks trained solely on protein sequences. The results confirm the effectiveness of the protein structure encoding in machine learning efforts on DMS data. Using structural representations as direct input to the networks, along with data augmentation and pretraining, significantly reduced demands on training data size and improved prediction performance, especially on smaller datasets, while performance on large datasets was on par with state-of-the-art sequence convolutional neural networks. The methods presented here have the potential to provide the same workflow as DMS without the experimental and financial burden of testing thousands of mutants. Additionally, we present an open-source, user-friendly software tool to make these data analysis techniques accessible, particularly to biotechnology and protein engineering researchers who wish to apply them to their mutagenesis data.