Find research datasets worth reusing
Search datasets from major research repositories and use ShareScore to quickly assess how well each record supports discovery, access, and reuse.
29
datasets available to search
ShareScore release 0.9.0
Dataset results
29 results for “Continual Learning”
Data: Learning Lattice Quantum Field Theories with Equivariant Continuous Flows
<p>Network parameters of continuous normalizing flows trained for the <span class="math-tex">\(\varphi^4\)</span> theory.</p> <p>Corresponding article: Learning Lattice Quantum Field Theories with Equivariant Continuous Flows [<a href="https://arxiv.org/abs/2207.00283">2207.00283</a>]</p> <p>Abstract: We propose a novel machine learning method for sampling from the high-dimensional probability distributions of Lattice Field Theories, which is based on a single neural ODE layer and incorporates the full symmetries of the problem. We test our model on the <span class="math-tex">\(\varphi^4\)</span> theory, showing that it systematically outperforms previously proposed flow-based methods in sampling efficiency, and the improvement is especially pronounced for larger lattices. Furthermore, we demonstrate that our model can learn a continuous family of theories at once, and the results of learning can be transferred to larger lattices. Such generalizations further accentuate the advantages of machine learning methods.</p>
Students' perceived obstacles with Forced Online Distance Learning during the CoVID-19 outbreak and their preferences to continue with the introduced teaching methods after the reopening of the University of Maribor [Project documentation]
<p>The outbreak of COVID -19 forced most universities into distance education. Three didacticians and researchers from the University of Maribor, Slovenia: Kosta Dolenc, Mateja Ploj Virtič and Andrej Šorgo formed a self-initiated initiative project group during the COVID -19 epidemic and started the first project with the working title: The Side Effects of Forced Online Distance Education (FODE).</p> <p>The aim of the second study, conducted during the first wave of the epidemic in March 2020, was to investigate the response of university students to the new situation. The project documentation provided for the Forced Online Distance Learning (FODL) consists of:</p> <ul> <li>abstract,</li> <li>instrument,</li> <li>copy of the descriptive statistics,</li> <li>and SPSS dataset.</li> </ul>
Learned value modulates the access to visual awareness during continuous flash suppression
<p>Data from Experiment 1 and Experiment 2 are reported in separate files. </p> <p>Each line contains the mean suppression time of a target grating under continuous flash suppression expressed in seconds for one participant. </p> <p>Each column refers to a different condition:<br> HREV = visual stimuli associated with high monetary reward<br> LREV = visual stimuli associated with low monetary reward<br> base = baseline measurements before associative learning<br> P1 = first measurement after associative learning<br> P2 = second measurement after associative learning<br> P3 = third measurement after associative learning</p> <p>For experiment 1, a short (20 trials) associative learning recall session was performed between P1 and P2 and between P2 and P3.</p> <p> </p> <p> </p> <p> </p>
HOWS-CL-25: Household Objects Within Simulation Dataset for Continual Learning
<p>HOWS-CL-25 (Household Objects Within Simulation dataset for Continual Learning) is a synthetic dataset especially designed for object classification on mobile robots operating in a changing environment (like a household), where it is important to learn new, never seen objects on the fly.<br> This dataset can also be used for other learning use-cases, like instance segmentation or depth estimation.<br> Or where household objects or continual learning are of interest.</p> <p>Our dataset contains 150,795 unique synthetic images using 25 different household categories with 925 3D models in total. For each of those categories, we generated about 6000 RGB images. In addition, we also provide a corresponding depth, segmentation, and normal image.</p> <p>The dataset was created with BlenderProc [Denninger et al. (2019)], a procedural pipeline to generate images for deep learning.<br> This tool created a virtual room with randomly textured floors, walls, and a light source with randomly chosen light intensity and color. After that, a 3D model is placed in the resulting room. This object gets customized by randomly assigning materials, including different textures, to achieve a diverse dataset. Moreover, each object might be deformed with a random<br> displacement texture.<br> We use 774 3D models from the ShapeNet dataset [A. X. Chang et al. (2015)] and the other models from various internet sites. Please note that we had to manually fix and filter most of the models with Blender before using them in the pipeline!</p> <p>For continual learning (CL), we provide two different loading schemes:<br> - Five sequences with five categories each<br> - Twelve sequences with three categories in the first and two in the other sequences.</p> <p>In addition to the RGB, depth, segmentation, and normal images, we also provide the calculated features of the RGB images (by ResNet50) as used in our RECALL paper.<br> In those two loading schemes, ten percent of the images are used for validation, where we ensure that an object instance is either in the training or the validation set, not in both. This avoids learning to recognize certain instances by heart.</p> <p>We recommend using those loading schemes to compare your approach with others.</p> <p>Here we provide three files for download:<br> - HOWS_CL_25.zip [124GB]: This is the original dataset with the RGB, depth, segmentation, and normal images, as well as the loading schemes. It is divided into three archive parts. To open the dataset, please ensure to download all three parts.<br> - HOWS_CL_25_hdf5_features.zip [2.5GB]: This only contains the calculated features from the RGB input by a ResNet50 in a .hdf5 file. Download this if you want to use the dataset for learning and/or want to compare your approach to our RECALL approach (where we used the same features).<br> - README.md: Some additional explanation.</p> <p>For further information and code examples, please have a look at our website: https://github.com/DLR-RM/RECALL.</p>
Solar Asset Mapper: A continuously-updated global inventory of solar energy facilities built with satellite data and machine learning
<p><strong>TransitionZero’s Solar Asset Mapper is a global, satellite-derived dataset of utility-scale solar farms generated with a combination of machine learning and human annotation. Our Q1 2024 dataset contains the location and shape of 63,616 assets, along with estimated capacities. We estimate the construction date for over 80% of these assets. The dataset contains over 19,100 square kilometres of solar farms across 183 countries, with a total estimated capacity of 711 GW.</strong></p> <p>Download the dataset, read the explainer and explore our polygon browser UI at <a href="https://www.transitionzero.org/products/solar-asset-mapper" target="_blank" rel="noopener">TransitionZero.org.</a></p> <p><a href="https://blog.transitionzero.org/hubfs/Data%20Products/TZ-SAM/tz-sam-scientific-methodology-Q12024.pdf" target="_blank" rel="noopener">Download our methodology paper here </a></p> <h1><strong>1. Dataset Description</strong></h1> <p>We publish six files.</p> <ul> <li><em>analysis_polygons.gpkg:</em> our “analysis-ready” dataset containing geometries, capacity estimates and construction date estimates.</li> <li><em>analysis_polygons.csv:</em> a version of analysis_polygons.gpkg containing a central latitude and longitude in place of a geometry, to allow parsing without geospatial software.</li> <li><em>sources.csv</em>: a table mapping the IDs of our analysis-ready dataset to the raw geometries that make them up.</li> <li><em>raw_polygons.gpkg:</em> the raw geometries used to compose analysis_polygons.gpkg.</li> <li><em>TZ Solar Asset Mapper Q1 2024.xlsx</em>: an Excel formatted version of the analysis_polygons.csv file.</li> <li><em>tz-sam_scientific_data.pdf</em>: A pre-print aricle that explains the methodology in detail.</li> </ul> <h2><strong>1.1 Analysis-level datasets</strong></h2> <p>Our analysis-level dataset comprises our most complete view of global asset-level solar installations, incorporating our own detections as well as known solar farm geometries from other datasets.</p> <p>The geospatial dataset contains the following fields:</p> <ul> <li>id: unique ID for the asset</li> <li>geometry: Polygon or MultiPolygon defining the asset</li> <li>capacity_mw: estimated capacity of the asset in megawatts</li> <li>constructed_before: upper bound for construction date (estimated date of the image in which the solar plant was first seen in a constructed state)</li> <li>constructed_after: lower bound for construction date (estimated date of the image in which construction began for the solar plant)</li> </ul> <p>The CSV version replaces the Geometry column with:</p> <ul> <li>latitude: the latitude of the centroid of the asset</li> <li>longitude: the longitude of the centroid of the asset</li> <li>country: administrative country name</li> </ul> <h2><strong>1.2 Raw datasets and sources</strong></h2> <p>The analysis-level datasets hide some complexity in the underlying data that we expose in the <em>raw_polygons</em> and <em>sources</em> file.</p> <ul> <li>We produce new sets of polygons for each run. Often these overlap, sometimes in complicated ways.</li> <li>We cluster together overlapping and nearby geometries from both our detections and external sources. Currently these sources are:</li> <li>Large solar farms scraped from OpenStreetMap (OSM)</li> <li>Validated geometries from <a href="../records/5005868">Kruitwagen et. al., A global inventory of solar photovoltaic generating units</a>.</li> </ul> <p>Each cluster comprises one row in the analysis-level dataset. In order to enable tracking raw detections from run to run, as well as to provide detailed sourcing information, we provide all of these raw polygons, along with a source file that lists all of the raw polygons contained in each analysis-level polygon.</p> <p>raw_polygons.gpkg contains the following fields:</p> <ul> <li>id: ID of the raw source polygon</li> <li>geometry<strong>: </strong>Polygon or MultiPolygon defining the asset</li> <li>source: either “solar asset mapper”, “osm” or “2019_global_pv”.</li> <li>acquisition_date: for solar asset mapper polygons, this is the date of the inference run that produced the polygon; for OSM polygons it is the date that the polygon was scraped from OSM; for 2019_global_pv it is 2019-01-01, the approximate detection date of that dataset.</li> </ul> <p>Sources.csv contains the following fields:</p> <ul> <li>cluster_id: ID of the corresponding item in the analysis-level dataset</li> <li>source_id: ID of the raw source polygon</li> <li>source: either “solar asset mapper”, “osm” or “2019_global_pv”.</li> <li>acquisition_date: for solar asset mapper polygons, this is the date of the inference run that produced the polygon; for OSM polygons it is the date that the polygon was scraped from OSM; for 2019_global_pv it is 2019-01-01, the approximate detection date of that dataset.</li> </ul> <h2><strong>1.3 Caveats and limitations</strong></h2> <h3><strong>1.3.1 Capacity Estimates</strong></h3> <p>While we have made every effort to remove false positives from the published dataset, some will remain due to the difficulty of manually validating detections in 10-metre satellite imagery. To estimate false positive prevalence throughout the data a subset of approximately 2000 detections were selected at random from our positively labelled solar assets. Each of these were validated through a higher degree of scrutiny utilising high-resolution imagery. This analysis yielded an expected rate of false positives of around 1%.</p> <h3>1.3.2 Plant Shapes</h3> <p>Our plant outlines are not perfect. They will occasionally be much smaller or larger than the underlying plant. Our tests show that on average, these effects average out.</p> <h3>1.3.3 Capacity Updates</h3> <p>Our capacity estimation model should produce relatively unbiased country-level aggregates, since it is trained to learn the typical ground coverage ratio of plants by country. The model has no way to distinguish between a very dense and a very sparse (e.g. dual-axis-tracking) plant in the same country. Plants with unusually high or low ground coverage ratios will not have accurate capacity estimates.</p> <h3>1.3.4 Construction Date Estimates</h3> <p>We are not able to directly estimate the construction date of a plant. We estimate an upper bound (the date of the image in which the plant was first seen in constructed state) and a lower bound (the date of the image in which the plant was last seen in an unconstructed state). For plants that were constructed before the launch date of Sentinel-2 in 2017, we produce only an upper bound.</p> <p>We leave it to consumers of the data to interpret these bounds and/or estimate likely grid connection dates.</p> <p><strong>2. Attribution</strong></p> <p>TZ-SAM is made available under a Creative Commons Attribution Non-Commercial 4.0 International License (CC-BY-NC-4.0). Attribution to TransitionZero is required. You must also clearly indicate if you have made any changes to the TZ-SAM dataset and what these are. Please refer to the suggested citation formats:</p> <ul> <li>“TransitionZero Solar Asset Mapper, TransitionZero, May 2024 release.”</li> <li>“TZ-SAM, TransitionZero, May 2024 release.”</li> <li>“TransitionZero (2024) Solar Asset Mapper.”</li> </ul>
Enhanced Westermo dataset - Transformed and Modified for Test case Selection and Priorotization in the context of Continuous Integration and Reinforcement Learning.
<p><strong>Overview</strong></p> <p>This repository contains a modified version of the existing, recently published dataset, Westermo. The initial dataset was gathered at Westermo Network Technologies AB, located in Västerås, Sweden. It encompasses over <strong>1 Million verdicts</strong> obtained from testing embedded systems, collected over a span of more than <strong>500 consecutive days</strong> of nightly testing. The dataset has been transformed and tailored specifically to cater to the research community, particularly for addressing challenges such as regression test selection, identification of flaky tests, and visualization of test results. The original dataset can be accessed through the reference provided in <strong>[1]</strong>.</p> <p>The Westermo dataset offers valuable historical information regarding the execution of test cases and their corresponding results. It serves as a valuable resource for evaluating and comparing different Test case Selection and Prioritization (TSP) techniques, enabling researchers to identify test cases that are more likely to fail during subsequent executions. Test cases in the dataset are characterized by attributes such as execution duration, previous last execution time, and the results of their recent executions.</p> <p>This dataset offers valuable historical information regarding the execution of test cases and their corresponding results. It serves as a valuable resource for evaluating and comparing different test case prioritization and selection techniques, enabling researchers to identify test cases that are more likely to fail during subsequent executions. Test cases in the dataset are characterized by attributes such as execution duration, previous last execution time, and the results of their recent executions.</p> <table align="left"> <caption><strong>Table 1: Dataset Overview</strong></caption> <tbody> <tr> <td>Test Cases</td> <td>1855</td> </tr> <tr> <td>CI Cycles</td> <td>15,197</td> </tr> <tr> <td>Verdict</td> <td>1,036,818</td> </tr> <tr> <td>Failed</td> <td>5.03%</td> </tr> </tbody> </table> <p> </p> <p> </p> <p> </p> <p> </p> <p> </p> <p> </p> <p> </p> <p>However, the diversity and multitude of the features in the dataset can be irrelevant to some TSP approaches. This led us to perform a dataset conversion, where we customized Westermo to have the same features from Paint Control and IOF/ROL, two widely used datasets in Reinforcement Learning based TSP approaches.</p> <p>This conversion required the combination of multiple variables and generating the target ones. When it comes to generating the “LastResults” and “Cycle” values, further analysis was required and the data handling needed an in-depth understanding of how the nightly testing was conducted. This led us to investigate what a CI cycle is in their context, and we followed their definition of a session, stating that “a session is when we run a suite of tests on one test system with a certain software version and testware version”. When splitting the data according to the 9 different systems used, we were able to generate 9 different sub-sets that fit the CI context.</p> <p> </p> <p><strong>File Format</strong></p> <p>The compressed .zip file contains 9 files, each one corresponding to each of the 9 systems. The datasets are available in CSV format, with the semicolon (;) serving as the delimiter. The columns included are represented in the table below along with their descriptions.</p> <table> <caption><strong>Table 2: Parameters of the dataset</strong></caption> <thead> <tr> <th scope="col">Column Name</th> <th scope="col">Content</th> </tr> </thead> <tbody> <tr> <td>Id</td> <td>Unique numeric identifier of the test execution </td> </tr> <tr> <td>Name</td> <td>Unique numeric identifier of the test case</td> </tr> <tr> <td>Duration</td> <td>Approximated runtime of the test case</td> </tr> <tr> <td>CalcPrio</td> <td>Priority of the test case, calculated by the prioritization algorithm (output column, initially 0)</td> </tr> <tr> <td>LastRun</td> <td>Previous last execution of the test case as date-time-string (Format: <em>YYYY-MM-DD HH:ii </em>)</td> </tr> <tr> <td>LastResults</td> <td>List of previous test results (Failed: 1, Passed: 0), ordered by ascending age. Lists are delimited by [ ].</td> </tr> <tr> <td>Verdict</td> <td> <p>Test verdict of this test execution (Failed: 1, Passed: 0)</p> </td> </tr> <tr> <td>Cycle</td> <td>The number of the CI cycle this test execution belongs to.</td> </tr> </tbody> </table> <p> </p> <p>The implications of this conversion are important as it can help the previous works to re-assess their approaches and have more data for training and testing, as well as opening a broader data spectrum for future researchers in this field to find ready-to-use, rich datasets, on which they could evaluate their approaches and contribute to the TSP community. This also addresses the limitations in the field discussed in the systematic literature review <strong>[2]</strong>, stating that future research on TSP techniques should focus on collecting data from more recent subjects in a CI context with varying failure rates and larger execution times, as reproducible studies with appropriate datasets are needed to develop a usable body of knowledge regarding TSP over time. We believe that this conversion of the Westermo dataset is our contribution to alleviating the gap for the RL-based approaches.</p> <p>The original dataset can be found <a href="https://sites.mdu.se/aidoart/results/open-source/test-results-dataset-westermo">here.</a></p>
Design of experiment (DOE) used in the study: Lightweight design of variable-stiffness imperfection-insensitive cylinders enabled by continuous tow shearing and machine learning
<p>There are five input variables that are changed for this design of experiment (DOE) within the following range:</p> <p><span class="math-tex">\(\begin{eqnarray} 0.05 \leq r_{CTS} \leq 0.20 \nonumber \\ 1 \leq n \leq 12 \nonumber \\ 0 \leq {c_2}_{ratio} \leq 1 \\ 0 \leq \theta_1 \leq 75 \nonumber \\ 0 \leq \theta_2 \leq 75 \nonumber \end{eqnarray}\)</span></p> <p>For the sake of simplicity, these variables are respectively called v1, v2, v3, v4, v5.</p> <p>The DOE consist of 2000 points created with Latin Hyper-cube Sampling, as available in the LHS toolbox (Carnell, R. lhs: Latin Hypercube Samples, 2021. R package version 1.1.3.).</p> <p>The outputs evaluated with this design of experiment (DOE) are the critical buckling load $P_{critical}$, $b_{factor}$ obtained with Koiter's asymptotic approach, and the mass.</p>
Learning to Do or Learning While Doing: Reinforcement Learning and Bayesian Optimisation for Online Continuous Tuning
<p>Dataset of optimisation runs performed for a study comparing reinforcement learning and Bayesian optimisation for online continuous tuning at the example of a linear particle accelerator tuning task.</p> <p> </p> <p><strong>Abstract of the Paper on the Study</strong></p> <p>Online tuning of real-world plants is a complex optimisation problem that continues to require manual intervention by experienced human operators. Autonomous tuning is a rapidly expanding field of research, where learning-based methods, such as Reinforcement Learning-trained Optimisation (RLO) and Bayesian optimisation (BO), hold great promise for achieving outstanding plant performance and reducing tuning times. Which algorithm to choose in different scenarios, however, remains an open question. Here we present a comparative study at the example of a routine task on a real particle accelerator, showing that RLO generally outperforms BO, but is not always the best choice. Based on the study’s results, we provide a clear set of criteria to guide the choice of algorithm for a given tuning task. These can ease the adoption of learning-based autonomous tuning solutions to the operation of complex real-world plants, ultimately improving the availability and pushing the limits of operability of these facilities, thereby enabling scientific and engineering<br> advancements.</p>
Dataset: Parameter estimation by learning quantum correlations in continuous photon-counting data using neural networks
<p>Dataset for the paper E. Rinaldi, M. González Lastre, S. García Herreros, S. Ahmed, M. Khanahmadi, F. Nori, and C. Sánchez Muñoz (2023), <a href="https://arxiv.org/abs/2310.02309">¨Parameter estimation by learning quantum correlations in continuous photon-counting data using neural networks¨, arxiv: 2310.02309</a></p> <p>This dataset can be used to populate the [datapath] folder in the repository <strong>ParamEst-NN</strong> (<a href="https://github.com/CarlosSMWolff/ParamEst-NN">github.com/CarlosSMWolff/ParamEst-NN</a> ) and reproduce the results shown in the paper.</p> <p>The dataset consist of four folders:</p> <ol> <li><strong>Training trajectories.</strong> Records of quantum-jump trajectories simulated with the Monte-Carlo solver of the <a href="https://qutip.org/">QuTiP</a> library, used to train neural networks for the problem of quantum parameter estimation. The records consist of time delays between quantum jumps.</li> <li><strong>Models. </strong>Models trained with the training trajectories provided, and used to obtain the results shown in the paper.</li> <li><strong>Validation trajectories. </strong>Trajectories used to benchmark the trained models. For the 2D case, the same trajectories are provided as a single .npy file, and split in 10 separated batches inside a ```batches``` folder. These are the batches that we used to generate Bayesian estimations on a cluster using nested sampling (see README file of the <a href="http://github.com/CarlosSMWolff/ParamEst-NN">repository</a>).</li> <li><strong>Cached results. </strong>Here we provide pre-computed Bayesian estimations for the 2D multi-parameter estimation case using nested sampling.</li> </ol> <p> </p> <p><em>E.R. was supported by Nippon Telegraph and Telephone Corporation (NTT) Research during the early stages of this work.<br> C.S.M. acknowledges that the project that gave rise to these results received the support of a fellowship from “la Caixa” Foundation (ID 100010434) and from the European Union’s Horizon 2020 research and innovation programme under the Marie Skłodowska-Curie Grant Agreement No.847648, with fellowship code LCF/BQ/PI20/11760026, and financial support from the MCINN project PID2021-126964OB-I00 (QENIGMA) and the Proyecto Sinérgico CAM 2020 Y2020/TCS- 6545 (NanoQuCo-CM).</em></p>
An Exploratory Evaluation of Continuous Feedback to Enhance Machine Learning Code Smell Detection
<p><em>Context</em>: Code smells are symptoms of bad design choices implemented on the source code. To manage and enhance software quality, it is important to be aware of code smells and refactor them whenever possible. As a result, several code smell detection tools and techniques have been proposed over the years. These tools and techniques present different strategies to detect code smells. More recently, machine learning algorithms have also been proposed to support code smell detection. However, we lack empirical evidence on how expert feedback could improve detection of these machine learning based techniques. <em>Objective:</em> This paper aims to propose and evaluate a machine-learning based strategy to improve detection of code smells by means of continuous feedback provided by the system expert.<em> Method: </em>To evaluate the strategy, we follow an experimental design to compare results of the detection before and after the feedback, both when feedback is provided at once and continuously. We focus on four code smells - God Class, Long Method, Feature Envy, and Refused Bequest - detected in twenty Java systems by using five code smell detection tools. We also extracted class- and method-level metrics from the systems for training the machine learning algorithms. <em>Results:</em> We observed that continuous feedback improves the performance of code smell detection. For the detection of God Class, a code smell with a detection performance initially good, we achieved an average improvement of 0.13 in terms of F1. For Refused Bequest, another class-level code smell, we achieved an average improvement of 0.58 in terms of F1 after all interactions of the strategy. For the method-level code smells, Long Method and Feature Envy, we achieved an average improvement of 0.66 and 0.72 in terms of F1, respectively.<em> Conclusions: </em>Our promising results are a stepping stone towards the development of tools relying on continuous feedback for machine learning detection of code smells.</p>
AbdomenCT-1K: Continual Learning Benchmark
<p>This is the dataset of AbdomenCT-1K: Continual Learning Benchmark.</p> <p>Related paper: <a href="https://ieeexplore.ieee.org/document/9497733/">https://ieeexplore.ieee.org/document/9497733/</a></p> <p>Benchmark homepage: https://abdomenct-1k-continual-learning.grand-challenge.org/</p>
Supplementary information for: A continuous-score occupancy modeling framework for incorporating uncertain machine learning output in autonomous biodiversity surveys
<p><span>Ecologists often study biodiversity by evaluating species occupancy and the relationship between occupancy and other covariates. Occupancy models are now widely used to account for false absences in field surveys and to reduce bias in estimates of covariate relationships. Existing occupancy models take as inputs binary detection/non-detection observations of species at each visit to each site. However, autonomous sensing devices and machine learning models are increasingly used to survey biodiversity, generating a new type of observation record (i.e., continuous-score data) that reflects the model's confidence a species is present in each autonomously sensed file, instead of binary detection/non-detection data. These data are not directly compatible with traditional binary occupancy modeling methods.</span></p> <p><span>Here, we develop a new occupancy model that models continuous scores on a visit level as a Gaussian mixture, combining a distribution of scores for files that do contain the species of interest and a distribution of scores for files that do not. The model takes as input continuous scores for each autonomously sensed and classified file, along with an optional small number of binary, manually verified detection and non-detection annotations.</span></p> <p><span>We present a simulation study that shows that over a range of empirically realistic parameters, our model outperforms traditional occupancy models that are based on binary annotation alone. We also apply this new model to an empirical case study using data generated from five machine learning classifiers applied to autonomous acoustic recordings gathered in the eastern United States.</span></p> <p><span>Because our occupancy model generalizes allowable input data beyond binary observations, it is particularly well-suited to the increasing volume of machine learning classified data in ecology and conservation.</span></p>
Dataset for TOPO-Loss for continuity-preserving crack detection using deep learning
<p>This is the dataset used to assess the performance of the crack detection algorithm proposed by Pantoja-Rosero et, al (2022) in the article "TOPO-Loss for continuity-preserving crack detection using deep learning" (https://doi.org/10.1016/j.conbuildmat.2022.128264)</p>
The reconstructed three-dimensional nitrate field dataset for the pan-European ocean using a continual learning-based multilayer perceptron
<p>Based on a newly developed continual learning-based multilayer perceptron model and environmental features, we reconstructed the pan-European 3D ocean nitrate field from 2010 to 2023. The reconstructed field features a monthly temporal resolution, a horizontal spatial resolution of 0.25 degrees, and 63 depth levels, with vertical intervals ranging from 5 to 50 m.</p>
CLEAR: Continual LEArning on Real-World Imagery
<p>Continual learning (CL) is considered as one of the next big challenges in AI. However, the existing CL benchmarks, e.g. Permuted-MNIST and Split-CIFAR, are artificially designed to be made continual and do not align with or generalize to real-world. In this paper, we introduce CLEAR, the first continual image recognition benchmark dataset with a natural temporal evolution of visual concepts in the real world that spans a decade (2004-2014). We build CLEAR from existing image collection (YFCC100M) by proposing a novel low-cost visio-linguistic dataset curation approach. It involves using pretrained vision-language models (e.g. CLIP) to quickly build high-quality labeled datasets on a tight budget. Finally, we post-process CLEAR via crowd-sourcing to remove errors and even inappropriate images hidden in original YFCC100M. The major strengths of CLEAR over prior CL benchmarks include (1) smooth and realistic temporal evolution of visual concepts with real-world imagery, enabling a more practical "online" (i.e., train on past, test on future) evaluation protocol (2) high-quality labeled data along with abundant unlabeled samples per time period for continual semi-supervised and unsupervised learning. Our extensive experiments reveal that mainstream "offline" evaluation protocols, which train and test on iid data, artificially inflate performance of CL systems, stressing the need for our "online" protocol since the models we train today will always be tested in future. Moreover, we find that state-of-the-arts CL algorithms that only utilize fully-supervised data fall short whereas unsupervised pretraining provides significant boost. Lastly, we introduce a biased reservoir-sampling algorithm that dynamically caches more recent training data, achieving the new state-of-the-arts while still leaving large room for improvement.</p>
Dynamic Knowledge Graphs for Continual Learning of Embeddings
<p>These datasets are generated from real world usecases. They are treated as Knowledge graphs and include 20 snapshots, where between two snapshots there are 10% added links and 10% deleted links, making the first and last snapshot non-overlapping.</p>
Supplementary information for: A continuous-score occupancy modeling framework for incorporating uncertain machine learning output in autonomous biodiversity surveys
Open the record for dataset details and reuse information.
Dataset and machine learning models for seismic response predictions of small-to-medium continuous girder bridges
<p>This upload includes the dataset and machine learning models (based on Matlab platform) for longitudinal seismic response predictions of multi-span highway girder bridges, which have a typical span length of 30 m supported by reinforced concrete (RC) bridge bents and abutments through spherical steel bearings. The input variables (features) are five structural parameters of studied bridges and seven intensity measures of earthquakes. The output variables (labels) are peak column drifts and peak bearing deformations. The dataset is developed by conducting a total number of 720 nonlinear time-history analyses considering the uncertainty of bridges and earthquakes. Machine learning models are developed using two popular machine learning algorithms named artificial neural network (ANN) and support vector regression (SVR).</p>
Dataset for Learning in Continuous Action Space for Developing High Dimensional Potential Energy Models
<p>The NN potentials developed in this study and the other available MLIP methods such as GAP, SNAP, qSNAP, and MEGNET used for benchmarking.</p>
Machine Learning Constructs Color Features to Accelerate Development of Long-Term Continuous Water Quality Monitoring
<p>This is a machine learning method for predicting the concentration of colored pollutants based on RGB and kmeans methods. This dataset includes raw images of pollutants as well as characteristic data of pollutants, as well as code for the model. You can see the contents of the zip file for details.</p>
ScienceDex guides
Understand access before you commit
These curated guides explain access requirements, typical timelines, costs, and reuse considerations for widely used research datasets.
Allen Brain Atlas
Allen Brain Atlas is an Allen Institute collection of brain map atlases, datasets, APIs, and analysis tools covering mouse, human, and non-human primate brain resources.
Annotated Behaviour and Observability Dataset (ABODe)
ABODe is a University of Edinburgh DataShare dataset for behavior classification in group-housed mice using home-cage video, identities, bounding boxes, ground-plate positions, and annotator labels.
DANDI Archive for NWB datasets
DANDI is a BRAIN Initiative archive for publishing and sharing neurophysiology data, including electrophysiology, optophysiology, and behavioral data packaged as NWB and related standards.
International Brain Laboratory public data
The International Brain Laboratory public data releases expose standardized mouse decision-making experiments, including Neuropixels recordings, widefield calcium imaging, behavior, and session metadata accessed through the ONE API.
OpenNeuro
OpenNeuro is a free, open platform for sharing neuroimaging datasets, with public search, dataset pages, and download paths for web, S3, DataLad, and the OpenNeuro CLI.