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.
1,048
datasets available to search
ShareScore release 0.9.0
Dataset results
1,048 results for “interpreting”
FIGURE 13 in Interpretation of anatomical characters in phylogenetic analysis of Pinnipedia, with emphasis on Otariidae (Mammalia, Carnivora)
FIGURE 13: Conspicuous metaconid in the fifth lower postcanine of Otaria byronia (MCN 2701 a, vestibular view; b, lingual view; scale, 5 cm).
FIGURE 12 in Interpretation of anatomical characters in phylogenetic analysis of Pinnipedia, with emphasis on Otariidae (Mammalia, Carnivora)
FIGURE 12: Multiple roots in the fifth upper postcanines (a, MCN 2624 Otaria byronia, vestibular view; b, MCN 2699 Arctocephalus australis, vestibular view; scale, 5 cm).
FIGURE 9 in Interpretation of anatomical characters in phylogenetic analysis of Pinnipedia, with emphasis on Otariidae (Mammalia, Carnivora)
FIGURE 9: Presence of the lingual cingulum in the third upper incisor of Arctocephalus australis (MCN 2509; scale, 5 cm).
FIGURE 6 in Interpretation of anatomical characters in phylogenetic analysis of Pinnipedia, with emphasis on Otariidae (Mammalia, Carnivora)
FIGURE 6: Condylar process above the level of the postcanine alveolar series in Otaria byronia (MCN 2629; scale, 5 cm).
FIGURE 2 in Interpretation of anatomical characters in phylogenetic analysis of Pinnipedia, with emphasis on Otariidae (Mammalia, Carnivora)
FIGURE 2: Orientation of the postcanine alveolar series showing medial concavity in Otaria byronia (MACN 4-3-12; scale, 2 cm).
FIGURE 1 in Interpretation of anatomical characters in phylogenetic analysis of Pinnipedia, with emphasis on Otariidae (Mammalia, Carnivora)
FIGURE 1: Nasal process of the premaxillary (arrow) not prominent in O. byronia (MCN 2629; scale, 2 cm).
FIGURE 11 in Interpretation of anatomical characters in phylogenetic analysis of Pinnipedia, with emphasis on Otariidae (Mammalia, Carnivora)
FIGURE 11: Multiple roots in the fourth upper postcanine of Arctocephalus australis (MCN 2456: a, vestibular view; b, apical view; scale, 5 cm).
FIGURE 8 in Interpretation of anatomical characters in phylogenetic analysis of Pinnipedia, with emphasis on Otariidae (Mammalia, Carnivora)
FIGURE 8: Internal auditory meatus showing incipient separation for the canals for the facial and vestibulocochlear nerves in Arctocephalus australis (MCN 2501; scale, 5 cm).
FIGURE 10 in Interpretation of anatomical characters in phylogenetic analysis of Pinnipedia, with emphasis on Otariidae (Mammalia, Carnivora)
FIGURE 10: Apical constriction and vertical sulcus on the root of the second upper postcanine of Arctocephalus australis (MCN 2456: a, lingual view; b, apical view; scale, 5 cm).
FIGURE 5 in Interpretation of anatomical characters in phylogenetic analysis of Pinnipedia, with emphasis on Otariidae (Mammalia, Carnivora)
FIGURE 5: Presence of a prolonged posterior-ventral projection in the wall of the tympanic bulla in Arctocephalus australis (MCN 1021; scale, 5 cm).
Interpretation of EKG with Image Recognition and Convolutional Neural Networks
<p>Dataset used in the training and evaluation of "Interpretation of EKG with Image Recognition and Convolutional Neural Networks"</p> <p>Abstract:</p> <p>Electrocardiograms (EKG) form the backbone of all cardiovascular diagnosis, treatment and follow up. Given the pivotal role it plays in modern medicine, there have been multiple efforts to computerize the EKG interpretation with algorithms to improve efficiency and accuracy. Unfortunately, many of these algorithms are machine specific and run-on proprietary signals generated by that machine, hence not generalizable. We propose the development of an image recognition model which can be used to read standard EKG strips. A convolutional neural network (CNN) was trained to classify 12-lead EKGs between 7 clinically important diagnostic classes. An austere variation of the MobileNetV3 model was trained from the ground up on publicly available labeled training set. The precision per class varies from 52% to 91%. This is a novel approach to EKG interpretation as an image recognition problem.</p>
Supporting data for "Reliable interpretability of biology-inspired deep neural networks"
<p><strong>Contents</strong></p> <p><em>data.tgz</em> contains all data necessary for reproducing the analysis in the manuscript. After cloning the GitHub repository, extract the contents of this file into folder <em>data</em>. The archive contains the following subfolders:</p> <ul> <li><em>dtox</em><br> DTox results, one subfolder per seed <ul> <li><em>module_relevance.tsv</em>: contains node importance scores, with the following columns: <ul> <li>(first, unnamed): compound identifier</li> <li>remaining columns: node identifiers (UniProt and Reactome IDs)</li> </ul> </li> <li><em>test_labels.csv</em>: predictions for the test set, with two columns: <ul> <li>truth: true label (0 or 1)</li> <li>predicted: predicted label (decimal number between 0 and 1)<br> </li> </ul> </li> </ul> </li> <li><em>mskimpact_[cancer type]_[experiment]</em><br> P-NET results using the MSK-IMPACT 2017 dataset, one subfolder per seed<br> [cancer type] is one of bc (breast cancer), cc (colorectal cancer), nsclc (non-small cell lung cancer), or pc (prostate cancer)<br> [experiment] is one of original (original setup) and shuffled (shuffled labels)<br> </li> <li><em>pnet_[experiment]</em><br> P-NET results using the original (prostate cancer) dataset, one subfolder per seed<br> [experiment] is one of deterministic (deterministic input data), original (original setup), and shuffled (shuffled labels) <ul> <li><em>node_importance.csv</em>: contains node importance scores, with the following columns: <ul> <li>(first, unnamed): node name</li> <li>coef: original node importance scores</li> <li>coef_graph: indegree plus outdegree of node</li> <li>coef_combined: adjusted node importance score (= coef / coef_graph if coef_graph > mean(coef_graph) + 5 sd(coef_graph) in the respective layer)</li> <li>coef_combined_zscore: scaled coef_combined</li> <li>coef_combined2: z(z(coef_graph) - z(coef))</li> <li>layer: layer of the node</li> </ul> </li> <li><em>predictions_test.csv</em>: predictions for the test set, with the following columns: <ul> <li>(first, unnamed): sample name</li> <li>pred: predicted class (unfortunately, encoded by a double 1.0 or 0.0)</li> <li>pred_scores: probability of the predicted class</li> <li>y: true class (encoded as integer 1 or 0)</li> </ul> </li> <li><em>predictions_train.csv</em>: predictions for the training set (same columns as above)</li> <li><em>link_weights_[layer].csv</em>: only in subfolder 234_20080808; matrices with edge weights</li> </ul> </li> </ul> <p> </p> <p><strong>Changelog</strong></p> <p><em>v1.1.0 – 2023-06-28</em></p> <ul> <li>added DTox results</li> <li>added results of P-NET experiments with MSK-IMPACT 2017 dataset</li> </ul> <p><em>v1.0.0 – 2023-03-22</em></p> <ul> <li>initial release</li> </ul>
ImageNet Mechanistic Interpretability
<p>To enable research on automated alignment/interpretability evaluations, we release the experimental results of our paper "Scale Alone Does not Improve Mechanistic Interpretability in Vision Models" as a separate dataset.</p> <p>Note that this is the first dataset containing interpretability measurements obtained through psychophysical experiments for multiple explanation methods and models. The dataset contains >120'000 anonymized human responses, each consisting of the final choice, a confidence score, and a reaction time. Out of these >120'000 responses, > 69'000 passed all our quality assertions - this is the main data (see responses_main.csv). The other responses failed (some) quality assertions and might be of lower quality - they should be used with care (see responses_lower_quality.csv). We consider the former the main dataset and provide the latter as data for development/debugging purposes. Furthermore, the dataset contains the used query images as well as the generated explanations for >760 units across nine models. </p> <p>The dataset itself is a collection of labels and metainformation without the presence of fixed features that should be predictive of a unit's interpretability. Moreover, finding and constructing features that are predictive of the recorded labels will be one of the open challenges posed by this line of research.</p>
BSL-Hansard: A parallel, multimodal corpus of English and interpreted British Sign Language data from parliamentary proceedings
<p>BSL-Hansard is a novel open source and multimodal resource composed by combining Sign Language video data in BSL and English text from the official transcription of British parliamentary sessions. This paper describes the method followed to compile BSL-Hansard including time alignment of text using the MAUS (Schiel, 2015) segmentation system, gives some statistics about this dataset, and suggests experiments. These primarily include end-to-end Sign Language-to-text translation, but is also relevant for broader machine translation, and speech and language processing tasks.</p> <p>This dataset will be useful for translation between BSL and English, or for studies in BSL or English down to the phonetic level.</p>
Sources of prey availability data alter interpretation of outputs from prey choice null networks
<p><em>Spider surveys</em></p> <p>Data collection was described previously by Cuff, Tercel, et al., (2022). This study pertains to a subset of those data, collected between 1<sup>st</sup> May and 9<sup>th</sup> July 2018 at 19 separate locations, for which paired sticky trap and vacuum sample data were collected (described below). Briefly, money spiders (Araneae: Linyphiidae) and wolf spiders (Araneae: Lycosidae) were visually located along transects in two adjacent barley fields at Burdons Farm, Wenvoe in South Wales (51°26'24.8"N, 3°16'17.9"W) and collected from webs and the ground. Transects were randomly distributed across the entire field. Along these transects, separate 4 m<sup>2</sup> quadrats, at least 10 m apart, were searched and all observed linyphiids and lycosids were collected. Spiders were placed in 100 % ethanol using an aspirator, regularly changing meshing to limit potential cross-contamination. Linyphiids occupying webs were prioritised for collection, but ground-active spiders were also collected. Spiders were taken to Cardiff University, transferred to fresh ethanol and stored at -80 °C in 100 % ethanol until DNA extraction. Extraction, amplification and sequencing of DNA, and bioinformatic analysis is described by Cuff, Tercel, et al., (2022) and Drake et al., (2022), and is also detailed below.</p> <p><em>Extraction and high-throughput sequencing of spider gut DNA</em></p> <p>Given their prevalence in field collections, dietary analysis was carried out for the linyphiid genera <em>Erigone</em>, <em>Tenuiphantes</em>, <em>Bathyphantes</em> and <em>Microlinyphia </em>(Araneae: Linyphiidae), and the Lycosidae genus <em>Pardosa</em>. Spiders were transferred to and washed in fresh 100 % ethanol to reduce external contaminants prior to identification via morphological key (Roberts, 1993). Abdomens were removed from spiders and again transferred to and washed in fresh 100 % ethanol. DNA was extracted from the abdomens via Qiagen TissueLyser II and DNeasy Blood & Tissue Kit (Qiagen) as per the manufacturer protocol, but with an extended lysis time of 12 hours to account for the complex and branched gut system in spider abdomens (Krehenwinkel et al., 2017).</p> <p>For amplification of DNA, two primer pairs were used. BerenF-LuthienR (Cuff et al., 2021) amplified a broad range of invertebrates including spiders, and TelperionF-LaureR (Cuff et al., 2022), amplified a range of invertebrates but fewer spiders. Primers were labelled with unique 10 bp molecular identifier tags (MID-tags) so that each individual had a unique pairing of forward and reverse tags for identification of each spider post-sequencing. PCR reactions of 25 µl contained 12.5 µl Qiagen PCR Multiplex kit, 0.2 µmol (2.5 µl of 2 µM) of each primer and 5 µl template DNA. Reactions were carried out in the same thermocycler, optimised via temperature gradient, with an initial 15 minutes at 95 °C, 35 cycles of 95 °C for 30 seconds, the primer-specific annealing temperature for 90 seconds and 72 °C for 90 seconds, respectively, followed by a final extension at 72 °C for 10 minutes. BerenF-LuthienR and TelperionF-LaureR used annealing temperatures of 52 °C and 42 °C, respectively.</p> <p>Within each PCR 96-well plate, 12 negative controls (extraction and PCR), 2 blank controls and 2 positive controls were included (i.e. 80 samples per plate), based on Taberlet <em>et al. </em>(2018). Positive controls were mixtures of invertebrate DNA comprised of non-native Asiatic species in four different proportions and blanks were empty wells within each plate to identify tag-jumping into unused MID-tag combinations. PCR negative controls were DNase-free water treated identically to DNA samples. A negative control was present for each MID-tag to identify any contamination of primers. All PCR products were visualised in a 2 % agarose gel with SYBRSafe (Thermo Fisher Scientific, Paisley, UK) and placed in categories based on their relative brightness. The concentration of these brightness categories was quantified via Qubit dsDNA High-sensitivity Assay Kits (Thermo Fisher Scientific, Waltham, MA, USA) with at least three representatives of each category per plate. The PCR products were then proportionally pooled according to these concentrations. Each pool was cleaned via SPRIselect beads (Beckman Coulter, Brea, USA), with a left-side size selection using a 1:1 ratio (retaining ~300-1000 bp fragments). The concentration of the pooled DNA was then determined via Qubit dsDNA High-sensitivity Assay Kits and pooled together into one library per primer pair. Library preparation for Illumina sequencing was carried out on the cleaned libraries via NEXTflex Rapid DNA-Seq Kit (Bioo Scientific, Austin, USA) and samples were sequenced on an Illumina MiSeq via a V3 chip with 300-bp paired-end reads (expected capacity ≤25,000,000 reads). Bioinformatic analysis followed Drake et al. (2022).</p> <p><em>Bioinformatic analysis</em></p> <p>The Illumina run generated 11,165,405 and 10,959,010 reads for BerenF-LuthienR and TelperionF-LaureR, respectively, which were quality-checked and paired via FastP (Chen et al., 2018) to retain only sequences of at least 200 bp with a quality threshold of 33, resulting in 10,561,874 and 9,355,112 paired reads. The paired reads were demultiplexed and assigned to their respective spider sample according to their MID-tags via the “trim.seqs” command in Mothur v1.39.5 (Schloss et al., 2009), leaving 7,854,610 and 7,437,929 reads with exact matches to the primer and MID-tags.</p> <p>Replicates were removed, and denoising and clustering to zero-radius operational taxonomic units (ZOTUs; clustered without % identity to avoid multiple species represented within a single operational taxonomic unit (OTU)) completed via Unoise3 in Usearch11 (Edgar, 2010). The resultant sequences were assigned a taxonomic identity from GenBank via BLASTn v2.7.1 (Camacho et al., 2009) using a 97 % identity threshold (Alberdi et al., 2017). The BLAST output was analysed in MEGAN v6.15.2 (Huson et al., 2016). Where the top BLAST hit, determined by lowest e-value, was resolved at a higher taxonomic level than species-level, the results were checked; where possibly erroneous entries were preventing species-level assignment (e.g., poorly resolved identifications on GenBank), finer resolution was assigned based on the next-closest match. Where ZOTUs were assigned the same taxon, these were aggregated.</p> <p>Data clean-up used the optimal minimum sequence copy thresholds identified by Drake et al. (2022). The maximum value for a ZOTU present in blank or negative controls was identified and subtracted from all read counts for that ZOTU to remove background contaminants. Simultaneously, known lab contaminants (e.g., German cockroach <em>Blattella germanica</em>), artefacts and errors of the sequencing process, unexpected reads in positive controls and positive control taxon reads in dietary samples were identified. These were calculated as a percentage of their respective sample’s read count and any read counts lower than the highest of these percentages for their respective sample were removed to eliminate additional instances of contamination. These thresholds were defined as 0.38 % and 0.39 % for BerenF-LuthienR and TelperionF-LaureR, respectively. The data from the two libraries (i.e., from each primer pair) were then aggregated together by sample and aggregated again by taxon. Non-target taxa (e.g., fungi) and instances in which predator DNA was amplified (i.e., ZOTUs with high read counts matching the individual’s morphological identity) were removed. </p> <p>The resultant sequencing read counts were converted into relative proportions (all values made to sum to one within each sample) and a mean value across the two primer pairs retained for each taxon within each sample. Relative read abundances were converted to presence-absence data of each detected prey taxon in each individual spider, but relative read abundance data were also retained for separate analyses to compare experimental outcomes between treatments.</p> <p><em>Invertebrate surveys</em></p> <p>To estimate prey availability using sticky traps, we placed one white dry 100 mm x 125 mm trap (Oecos) in the 4 m<sup>2</sup> quadrat centred at the position where the spider was captured. The trap was suspended with wire approximately 25 mm above the ground to catch falling, crawling and flying invertebrates, and left in place for 72 hours. Invertebrates were identified on the traps under a stereomicroscope. To estimate prey availability using suction sampling, ground and crop stems were sampled using a ‘G-vac’ for approximately 30 seconds at each location. The collected material was emptied into a bag, any organisms immediately killed with ethyl-acetate and material frozen for storage before sorting into 70 % ethanol in the lab. All invertebrates were identified to family level to match the resolution of the least resolved of the metabarcoding-derived trophic interaction data, and due to difficulties associated with identification to finer taxonomic resolution for many taxa. Exceptions included springtails of the superfamily Sminthuroidea (Sminthuridae and Bourletiellidae were often indistinguishable following suction sampling and preservation due to the fine features necessary to distinguish them) which were left at super-family, mites (many of which were immature or in poor condition) which were identified to order level, and wasps of the superfamily Ichneumonoidea which were identified no further due to obscurity of wing venation due to damage following suction sampling.</p> <p><em>Statistical Analysis</em></p> <p>All analyses were conducted in R v4.0.3 (R Core Team, 2021) and carried out on invertebrate data at the family or superfamily level. Alongside the dietary data derived from metabarcoding, and prey availability as determined directly by suction sampling (abundance) and sticky trapping (activity density), three additional datasets were generated where two were designed to combine data from the two trapping methods. The first approach simply set all invertebrate taxa detected in the field to have equal abundance, to provide a baseline against which to assess the effects of different prey abundance estimates. When generating the two combined data sets, it was apparent that simply adding them together would underrepresent one of the datasets as abundance and activity density are measured in different units. Therefore, a ‘proportional combined’ dataset was generated by converting counts to relative proportions of each sample (to equally weight the two methods), which were then combined by summing proportions between the two methods for each sample, multiplied by the total count of individuals across both methods for each sample (to create realistic abundance values), and then rounded to the nearest integer (to return count data). In addition, a ‘frequency of occurrence (FOO) combined’ dataset was generated by converting counts to binary presence-absence values of each sample, which were then summed between the two methods for each sample. To assess the diversity represented by the two sampling methods and their combinations, and the completeness of those datasets, coverage-based rarefaction and extrapolation were carried out, and Hill diversity calculated (Chao et al., 2014; Roswell et al., 2021) using the ‘iNEXT’ package with families represented by frequency-of-occurrence across samples (Chao et al., 2014; Hsieh et al., 2016).</p> <p>The remaining analyses were performed using both presence-absence and relative read abundance dietary data separately to show how differences in the treatment of the observed data are reflected in the outcomes of the analyses. Figures and outputs given in the main text relate to the presence-absence data, while relative read abundance figures and outputs are presented in the Supplementary Information. Prey preferences of spiders were analysed using network-based null models in the ‘econullnetr’ package (Vaughan et al., 2018) with the ‘generate_null_net’ function. Econullnetr generates null models based on prey availability to predict how consumers would forage if based on the availability of resources alone. These null models are then compared against the observed interactions of consumers (e.g., interactions of spiders with their prey based on dietary metabarcoding) to ascertain the extent to which resource consumption deviated from random. In five separate null models, prey availability was represented separately by the datasets described above: abundance (suction sampling), activity density (sticky trapping), proportional combined, FOO combined and equal prey abundance.</p> <p>To compare effect sizes between null models for each resource taxon, mean prey preference standardised effect size (SES) values were calculated from the individual spiders per model. The SES values were plotted and joined between taxa to visualise paired differences using ‘ggplot’ (Wickham, 2016). Null model-predicted trophic interactions were generated via an econullnetr null model with 999 simulations with outputs extended to allow the comparison of the null interactions for individual consumers (generate_null_net_indiv; Cuff, Kitson, et al., 2023). A visualisation of the per-individual differences in null model and observed data was generated via non-metric multi-dimensional scaling (NMDS) using the ‘metaMDS’ function in the ‘vegan’ package (Oksanen et al., 2016) in two dimensions and 9999 simulations, with Euclidean distance. Centroid coordinates for each null model and the observed data were extracted and pairwise distances calculated between model centroids:</p> <p>The ‘observed’ network (i.e., the network determined solely by dietary data, not necessarily the objectively ‘true’ network) and each null network were visualised with the associated prey choice effect sizes as a bipartite network using ‘ggnetwork’ (Briatte, 2021; Wickham, 2016) via an ‘igraph’ object (Csardi & Nepusz, 2006). The degree of each prey node, weighted nestedness and linkage density were generated using the ‘bipartite’ package (Dormann et al., 2008) for each network and compared visually via ggplot2.</p>
Replication Package for "Improving the Readability of Generated Tests Using GPT-4 and ChatGPT Code Interpreter"
<p>While automated test generation can decrease the human burden associated with testing, it does not eliminate this burden. Humans must still work with generated test cases to interpret testing results, debug the code, build and maintain a comprehensive test suite, and many other tasks. Therefore, a major challenge with automated test generation is understandability of generated test test cases. </p> <p>Large language models (LLMs), machine learning models trained on massive corpora of textual data - including both natural language and programming languages - are an emerging technology with great potential for performing language-related predictive tasks such as translation, summarization, and decision support. </p> <p>In this study, we are exploring the capabilities of LLMs with regard to improving test case understandability.</p> <p>This package contains the data produced during this exploration:</p> <ul> <li>The examples directory contains the three case studies we tested our transformation process on: <ul> <li>queue_example: Tests of a basic queue data structure</li> <li>httpie_sessions: Tests of the sessions module from the httpie project. </li> <li>string_utils_validation: Tests of the validation module from the python-string-utils project.</li> <li>Each directory contains the modules-under-test, the original test cases generated by Pynguin, and the transformed test cases. </li> <li>Two trials were performed per case example of the transformation technique to assess the impact of different results from the LLM.</li> </ul> </li> <li>The survey directory contains the survey that was sent to assess the impact of the transformation on test readability. <ul> <li>survey.pdf contains the survey questions.</li> <li>responses.xlsx contains the survey results.</li> </ul> </li> </ul>
Interpretation results used in the paper "Real-Time Dual-Para FWI-Inv of GPR Data Based on Robust Deep Learning"
<p>We provide the updated synthetic data and inversion results used for all interpretations presented in the paper "Real-Time Dual-Parameter FWI Inversion of GPR Data Based on Robust Deep Learning."</p>
Data for: Assessing the Cognition of Movement Trajectory Visualizations: Interpreting Speed and Direction
Open the record for dataset details and reuse information.
The western United States large forest-fire stochastic simulator (WULFFSS) 1.0: A monthly gridded forest-fire model using interpretable statistics
Open the record for dataset details and reuse information.
Data from: Cornerstones are the key stones: Using interpretable machine learning to probe the clogging process in 2D granular hoppers
Open the record for dataset details and reuse information.
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.