Thierry Artières
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Previous Organizations / Institutes
The organizations / institutes from the scholar's other articles in the past ten years
Thierry Artières has a total of 113 co-authors. Here are the top co-authors by number of shared publications:
P. Gallinari
8 shared publicationsPatrick Gallinari
Sorbonne University, Paris, France
6 shared publications- Sanparith Marukatat
National Electronics and Computer Technology Center (NECTEC)
6 shared publications B. Dorizzi
EPH, Institut National des Télécommunications
5 shared publications- Sylvain Takerkart
Aix-Marseille University, CNRS, INT, UMR7289 1 , F-13005 Marseille
5 shared publications - Stéphane Ayache
Laboratoire d’Informatique et Systèmes UMR 7020, CNRS, Aix-Marseille University
4 shared publications
Latest Publications
- Open Access
- 13 October 2026
- Published by EDP Sciences in Astronomy & Astrophysics
- Vol. 714, A208
Deep learning for filament segmentation: Hyperparameter estimation and limitations
Context . Detecting filamentary structures in molecular clouds is essential for understanding star formation, yet remains difficult due to the evolving nature of filaments, their wide variety, and in practice incomplete datasets (partially labeled data). Despite promising results with deep learning, the filament detection field lacks standardized, ground-truth–complete benchmarks for method comparison leading to method improvement. Aims . We use a synthetic benchmark dataset to study how filament detection methods behave when their parameters are selected automatically from training and validation sets. The objective is to test the scalability of a single, globally optimized configuration (hyper-parameter set) to a large and heterogeneous dataset. We also assess the impact of incomplete annotations – common in observational datasets – on model bias, focusing on low-contrast and low-density filaments. Methods . We introduce a synthetic benchmark based on Plummer-like radial profiles with tunable parameters and spatially coherent backgrounds. This is motivated by observational results of filament properties (Herschel PACS/SPIRE). The synthetic dataset yields realistic column-density maps with fully known ground truth. We leverage this dataset to apply a common training-validation-test protocol to three approaches: UNet for semantic segmentation, Mask R-CNN for instance segmentation, and FilFinder as a morphology-based method. For FilFinder , the experiment should be understood as an automatic grid-search study over global parameters, rather than as an assessment of the method under manual, field-specific tuning. Results . Under this global automatic-tuning protocol, FilFinder is difficult to summarize with a single parameter set across the full heterogeneity of the synthetic dataset. In contrast, UNet and Mask-RCNN adapt their internal parameters during training and achieve stronger aggregate scores. Semantic and instance segmentation paradigms exhibit complementary strengths. Crucially, training on observationally biased labels induces pronounced failures on low-contrast (low- p ) and low-density filaments; unbiased training restores performance across these regimes. Conclusions . The proposed Plummer-based benchmark provides a rigorous and flexible framework for studying both automatic parameter tuning and learning biases in filament detection. More broadly, the results emphasize the need for complete, physically informed supervision and caution against training on incomplete catalogs. Synthetic, physics-aware data therefore provide a useful controlled environment for method development, parameter-selection studies, and bias diagnosis.
- 13 October 2026
- Published by EDP Sciences in Astronomy & Astrophysics
- Vol. 714, A208
Open AccessDeep learning for filament segmentation: Hyperparameter estimation and limitations
Context . Detecting filamentary structures in molecular clouds is essential for understanding star formation, yet remains difficult due to the evolving nature of filaments, their wide variety, and in practice incomplete datasets (partially labeled data). Despite promising results with deep learning, the filament detection field lacks standardized, ground-truth–complete benchmarks for method comparison leading to method improvement. Aims . We use a synthetic benchmark dataset to study how filament detection methods behave when their parameters are selected automatically from training and validation sets. The objective is to test the scalability of a single, globally optimized configuration (hyper-parameter set) to a large and heterogeneous dataset. We also assess the impact of incomplete annotations – common in observational datasets – on model bias, focusing on low-contrast and low-density filaments. Methods . We introduce a synthetic benchmark based on Plummer-like radial profiles with tunable parameters and spatially coherent backgrounds. This is motivated by observational results of filament properties (Herschel PACS/SPIRE). The synthetic dataset yields realistic column-density maps with fully known ground truth. We leverage this dataset to apply a common training-validation-test protocol to three approaches: UNet for semantic segmentation, Mask R-CNN for instance segmentation, and FilFinder as a morphology-based method. For FilFinder , the experiment should be understood as an automatic grid-search study over global parameters, rather than as an assessment of the method under manual, field-specific tuning. Results . Under this global automatic-tuning protocol, FilFinder is difficult to summarize with a single parameter set across the full heterogeneity of the synthetic dataset. In contrast, UNet and Mask-RCNN adapt their internal parameters during training and achieve stronger aggregate scores. Semantic and instance segmentation paradigms exhibit complementary strengths. Crucially, training on observationally biased labels induces pronounced failures on low-contrast (low- p ) and low-density filaments; unbiased training restores performance across these regimes. Conclusions . The proposed Plummer-based benchmark provides a rigorous and flexible framework for studying both automatic parameter tuning and learning biases in filament detection. More broadly, the results emphasize the need for complete, physically informed supervision and caution against training on incomplete catalogs. Synthetic, physics-aware data therefore provide a useful controlled environment for method development, parameter-selection studies, and bias diagnosis.
- Open Access
- 8 January 2026
- Published by eLife Sciences Publications, Ltd in eLife
- Vol. 13
Reconstructing voice identity from noninvasive auditory cortex recordings
The cerebral processing of voice information is known to engage, in human as well as non-human primates, ‘temporal voice areas’ (TVAs) that respond preferentially to conspecific vocalizations. However, how voice information is represented by neuronal populations in these areas, particularly speaker identity information, remains poorly understood. Here, we used a deep neural network (DNN) to generate a high-level, small-dimension representational space for voice identity—the ‘voice latent space’ (VLS)—and examined its linear relation with cerebral activity via encoding, representational similarity, and decoding analyses. We find that the VLS maps onto fMRI measures of cerebral activity in response to tens of thousands of voice stimuli from hundreds of different speaker identities and better accounts for the representational geometry for speaker identity in the TVAs than in A1. Moreover, the VLS allowed TVA-based reconstructions of voice stimuli that preserved essential aspects of speaker identity as assessed by both machine classifiers and human listeners. These results indicate that the DNN-derived VLS provides high-level representations of voice identity information in the TVAs.
- 8 January 2026
- Published by eLife Sciences Publications, Ltd in eLife
- Vol. 13
Open AccessReconstructing voice identity from noninvasive auditory cortex recordings
The cerebral processing of voice information is known to engage, in human as well as non-human primates, ‘temporal voice areas’ (TVAs) that respond preferentially to conspecific vocalizations. However, how voice information is represented by neuronal populations in these areas, particularly speaker identity information, remains poorly understood. Here, we used a deep neural network (DNN) to generate a high-level, small-dimension representational space for voice identity—the ‘voice latent space’ (VLS)—and examined its linear relation with cerebral activity via encoding, representational similarity, and decoding analyses. We find that the VLS maps onto fMRI measures of cerebral activity in response to tens of thousands of voice stimuli from hundreds of different speaker identities and better accounts for the representational geometry for speaker identity in the TVAs than in A1. Moreover, the VLS allowed TVA-based reconstructions of voice stimuli that preserved essential aspects of speaker identity as assessed by both machine classifiers and human listeners. These results indicate that the DNN-derived VLS provides high-level representations of voice identity information in the TVAs.
- Open Access
- 20 November 2025
- Published by eLife Sciences Publications, Ltd in eLife preprints
Reconstructing Voice Identity from Noninvasive Auditory Cortex Recordings
The cerebral processing of voice information is known to engage, in human as well as non-human primates, “temporal voice areas” (TVAs) that respond preferentially to conspecific vocalizations. However, how voice information is represented by neuronal populations in these areas, particularly speaker identity information, remains poorly understood. Here, we used a deep neural network (DNN) to generate a high-level, small-dimension representational space for voice identity—the ‘voice latent space’ (VLS)—and examined its linear relation with cerebral activity via encoding, representational similarity, and decoding analyses. We find that the VLS maps onto fMRI measures of cerebral activity in response to tens of thousands of voice stimuli from hundreds of different speaker identities and better accounts for the representational geometry for speaker identity in the TVAs than in A1. Moreover, the VLS allowed TVA-based reconstructions of voice stimuli that preserved essential aspects of speaker identity as assessed by both machine classifiers and human listeners. These results indicate that the DNN-derived VLS provides high-level representations of voice identity information in the TVAs.
- 20 November 2025
- Published by eLife Sciences Publications, Ltd in eLife preprints
Open AccessReconstructing Voice Identity from Noninvasive Auditory Cortex Recordings
The cerebral processing of voice information is known to engage, in human as well as non-human primates, “temporal voice areas” (TVAs) that respond preferentially to conspecific vocalizations. However, how voice information is represented by neuronal populations in these areas, particularly speaker identity information, remains poorly understood. Here, we used a deep neural network (DNN) to generate a high-level, small-dimension representational space for voice identity—the ‘voice latent space’ (VLS)—and examined its linear relation with cerebral activity via encoding, representational similarity, and decoding analyses. We find that the VLS maps onto fMRI measures of cerebral activity in response to tens of thousands of voice stimuli from hundreds of different speaker identities and better accounts for the representational geometry for speaker identity in the TVAs than in A1. Moreover, the VLS allowed TVA-based reconstructions of voice stimuli that preserved essential aspects of speaker identity as assessed by both machine classifiers and human listeners. These results indicate that the DNN-derived VLS provides high-level representations of voice identity information in the TVAs.
- Open Access
- 22 June 2025
- Published by Wiley in NMR in Biomedicine
- Vol. 38 (8) , e70079
Insights on Scan‐Specific Deep‐Learning Strategies for Brain MRI Parallel Imaging Reconstruction
Scan‐specific deep learning strategies have been proposed for parallel imaging reconstruction in which auto‐calibrated signals (ACS) are used for training. Here, we introduce methods to objectively optimize architecture and training details. In addition, we introduce a new metric to better characterize the quality of the reconstructed images. Various brain MRI situations are considered. The evaluated models encompass single‐layer and three‐layer residual CNN architectures with real and complex convolutions. Hyperparameters such as the level of linearity in leaky activation functions, loss function, kernel sizes and depths are optimized using grid‐search with K‐fold cross validation. The performances regarding ACS reference size and mode are also evaluated. An innovative COrrelation‐Based Residual Artifact Index (COBRAI) quantifying the level of structured residual artifacts is proposed. Qualitative and quantitative comparisons are performed both on the FastMRI and in‐house multi‐contrasts 2D data. The proposed objective grid‐search strategy based on ACS successfully provided optimized hyperparameters, retrospectively validated by enhanced image quality metrics. Notably, it is shown that nonlinearities produce structured residual artifacts, and that, among the models tested, a three‐layer residual linear CNN with complex implementation and a reduced number of parameters is more robust, particularly providing less structured artifacts with less training data, leading to larger acceleration rates. Deep‐learning MRI parallel image reconstruction in the scan‐specific approach can be optimized using grid‐search with K‐fold cross validation. It was successfully applied in various 2D brain MRI situations. The quantification of structured residual artifacts with COBRAI is a useful complementary characterization to state‐of‐the‐art metrics, and it can be used to drive model selection.
- 22 June 2025
- Published by Wiley in NMR in Biomedicine
- Vol. 38 (8) , e70079
Open AccessInsights on Scan‐Specific Deep‐Learning Strategies for Brain MRI Parallel Imaging Reconstruction
Scan‐specific deep learning strategies have been proposed for parallel imaging reconstruction in which auto‐calibrated signals (ACS) are used for training. Here, we introduce methods to objectively optimize architecture and training details. In addition, we introduce a new metric to better characterize the quality of the reconstructed images. Various brain MRI situations are considered. The evaluated models encompass single‐layer and three‐layer residual CNN architectures with real and complex convolutions. Hyperparameters such as the level of linearity in leaky activation functions, loss function, kernel sizes and depths are optimized using grid‐search with K‐fold cross validation. The performances regarding ACS reference size and mode are also evaluated. An innovative COrrelation‐Based Residual Artifact Index (COBRAI) quantifying the level of structured residual artifacts is proposed. Qualitative and quantitative comparisons are performed both on the FastMRI and in‐house multi‐contrasts 2D data. The proposed objective grid‐search strategy based on ACS successfully provided optimized hyperparameters, retrospectively validated by enhanced image quality metrics. Notably, it is shown that nonlinearities produce structured residual artifacts, and that, among the models tested, a three‐layer residual linear CNN with complex implementation and a reduced number of parameters is more robust, particularly providing less structured artifacts with less training data, leading to larger acceleration rates. Deep‐learning MRI parallel image reconstruction in the scan‐specific approach can be optimized using grid‐search with K‐fold cross validation. It was successfully applied in various 2D brain MRI situations. The quantification of structured residual artifacts with COBRAI is a useful complementary characterization to state‐of‐the‐art metrics, and it can be used to drive model selection.
- Open Access
- 3 December 2024
- Published by EDP Sciences in Astronomy & Astrophysics
- Vol. 692, A41
Supervised machine learning on Galactic filaments
Context. Filaments host star formation and are fundamental structures of galaxies. Their diversity, as observed in the interstellar medium, from very low-density structures to very dense hubs, and their complex life cycles make their complete detection challenging over this large diversity range. Aims. Using 2D H2 column density images obtained as part of the Herschel Hi-GAL survey of the Galactic plane (Gp), we want to detect, simultaneously and using a single model, filaments over a large range of column density and contrast over the whole Gp. In particular, we target low-contrast and low-density structures that are particularly difficult to detect with classical algorithms. Methods. The whole H2 column density image of the Gp was subdivided into individual patches of 32 × 32 pixels. Following our proof of concept study aimed at exploring the potential of supervised learning for the detection of filaments, we propose an innovative supervised learning method based on adding information by encoding the position of these patches in the Gp. To allow the segmentation of the whole Gp, we introduced a random procedure that preserves the balance within the model training and testing datasets over the Gp plane. Four architectures and six models were tested and compared using different metrics. Results. For the first time, a segmentation of the whole Gp has been obtained using supervised deep learning. A comparison of the models based on metrics and astrophysical results shows that one of the architectures (PE-UNet-Latent), where the position encoding was done in the latent space gives the best performance to detect filaments over the whole range of density and contrast observed in the Gp. A normalized map of the whole Gp was also produced and reveals the highly filamentary structure of the Gp in all density regimes. We successfully tested the generalization of our best model by applying it to the 2D 12CO COHRS molecular data obtained on a 58.°8 portion (in longitude) of the plane. Conclusions. We demonstrate the interest of position encoding to allow the detection of filaments over the wide range of density and contrast observed in the Gp. The produced maps (both normalized and segmented) offer a unique opportunity for follow-up studies of the life cycle of Galactic filaments. The promising generalization possibility tested on a molecular dataset of the Gp opens new opportunities for systematic detection of filamentary structures in the big data context available for the Gp.
- 3 December 2024
- Published by EDP Sciences in Astronomy & Astrophysics
- Vol. 692, A41
Open AccessSupervised machine learning on Galactic filaments
Context. Filaments host star formation and are fundamental structures of galaxies. Their diversity, as observed in the interstellar medium, from very low-density structures to very dense hubs, and their complex life cycles make their complete detection challenging over this large diversity range. Aims. Using 2D H2 column density images obtained as part of the Herschel Hi-GAL survey of the Galactic plane (Gp), we want to detect, simultaneously and using a single model, filaments over a large range of column density and contrast over the whole Gp. In particular, we target low-contrast and low-density structures that are particularly difficult to detect with classical algorithms. Methods. The whole H2 column density image of the Gp was subdivided into individual patches of 32 × 32 pixels. Following our proof of concept study aimed at exploring the potential of supervised learning for the detection of filaments, we propose an innovative supervised learning method based on adding information by encoding the position of these patches in the Gp. To allow the segmentation of the whole Gp, we introduced a random procedure that preserves the balance within the model training and testing datasets over the Gp plane. Four architectures and six models were tested and compared using different metrics. Results. For the first time, a segmentation of the whole Gp has been obtained using supervised deep learning. A comparison of the models based on metrics and astrophysical results shows that one of the architectures (PE-UNet-Latent), where the position encoding was done in the latent space gives the best performance to detect filaments over the whole range of density and contrast observed in the Gp. A normalized map of the whole Gp was also produced and reveals the highly filamentary structure of the Gp in all density regimes. We successfully tested the generalization of our best model by applying it to the 2D 12CO COHRS molecular data obtained on a 58.°8 portion (in longitude) of the plane. Conclusions. We demonstrate the interest of position encoding to allow the detection of filaments over the wide range of density and contrast observed in the Gp. The produced maps (both normalized and segmented) offer a unique opportunity for follow-up studies of the life cycle of Galactic filaments. The promising generalization possibility tested on a molecular dataset of the Gp opens new opportunities for systematic detection of filamentary structures in the big data context available for the Gp.
- Open Access
- 5 April 2024
- Published by Public Library of Science (PLoS) in PLOS Computational Biology
- Vol. 20 (4) , e1012006
Partial label learning for automated classification of single-cell transcriptomic profiles
Single-cell RNA sequencing (scRNASeq) data plays a major role in advancing our understanding of developmental biology. An important current question is how to classify transcriptomic profiles obtained from scRNASeq experiments into the various cell types and identify the lineage relationship for individual cells. Because of the fast accumulation of datasets and the high dimensionality of the data, it has become challenging to explore and annotate single-cell transcriptomic profiles by hand. To overcome this challenge, automated classification methods are needed. Classical approaches rely on supervised training datasets. However, due to the difficulty of obtaining data annotated at single-cell resolution, we propose instead to take advantage of partial annotations. The partial label learning framework assumes that we can obtain a set of candidate labels containing the correct one for each data point, a simpler setting than requiring a fully supervised training dataset. We study and extend when needed state-of-the-art multi-class classification methods, such as SVM, kNN, prototype-based, logistic regression and ensemble methods, to the partial label learning framework. Moreover, we study the effect of incorporating the structure of the label set into the methods. We focus particularly on the hierarchical structure of the labels, as commonly observed in developmental processes. We show, on simulated and real datasets, that these extensions enable to learn from partially labeled data, and perform predictions with high accuracy, particularly with a nonlinear prototype-based method. We demonstrate that the performances of our methods trained with partially annotated data reach the same performance as fully supervised data. Finally, we study the level of uncertainty present in the partially annotated data, and derive some prescriptive results on the effect of this uncertainty on the accuracy of the partial label learning methods. Overall our findings show how hierarchical and non-hierarchical partial label learning strategies can help solve the problem of automated classification of single-cell transcriptomic profiles, interestingly these methods rely on a much less stringent type of annotated datasets compared to fully supervised learning methods. Recent years have witnessed an exponential increase in the amount of single-cell RNASeq data generated, particularly in studies of development. One of the major challenges is to identify individual cell types within the data. Expert knowledge is required to identify the relevant marker genes, tissue and timing that will enable the cell type identification. This information can be difficult to obtain and calls for automated cell type classification approaches. Classical classification techniques would solve this problem by training a classifier on a fully supervised dataset. However, this only pushes the problem further, as a dataset annotated at single-cell resolution is still needed for training. Here we propose instead to take advantage of the partial label learning framework which let us train our classifiers on a set of candidate labels per transcriptomic profile. This approach overcomes the need for a training dataset annotated at single-cell resolution. We show that we obtain classification accuracy similar to the fully supervised case. We explore the effect of varying the amount of partially labeled data and of considering the hierarchical structure of the label set (derived from the developmental processes) in the models on simulated and real biological datasets.
- 5 April 2024
- Published by Public Library of Science (PLoS) in PLOS Computational Biology
- Vol. 20 (4) , e1012006
Open AccessPartial label learning for automated classification of single-cell transcriptomic profiles
Single-cell RNA sequencing (scRNASeq) data plays a major role in advancing our understanding of developmental biology. An important current question is how to classify transcriptomic profiles obtained from scRNASeq experiments into the various cell types and identify the lineage relationship for individual cells. Because of the fast accumulation of datasets and the high dimensionality of the data, it has become challenging to explore and annotate single-cell transcriptomic profiles by hand. To overcome this challenge, automated classification methods are needed. Classical approaches rely on supervised training datasets. However, due to the difficulty of obtaining data annotated at single-cell resolution, we propose instead to take advantage of partial annotations. The partial label learning framework assumes that we can obtain a set of candidate labels containing the correct one for each data point, a simpler setting than requiring a fully supervised training dataset. We study and extend when needed state-of-the-art multi-class classification methods, such as SVM, kNN, prototype-based, logistic regression and ensemble methods, to the partial label learning framework. Moreover, we study the effect of incorporating the structure of the label set into the methods. We focus particularly on the hierarchical structure of the labels, as commonly observed in developmental processes. We show, on simulated and real datasets, that these extensions enable to learn from partially labeled data, and perform predictions with high accuracy, particularly with a nonlinear prototype-based method. We demonstrate that the performances of our methods trained with partially annotated data reach the same performance as fully supervised data. Finally, we study the level of uncertainty present in the partially annotated data, and derive some prescriptive results on the effect of this uncertainty on the accuracy of the partial label learning methods. Overall our findings show how hierarchical and non-hierarchical partial label learning strategies can help solve the problem of automated classification of single-cell transcriptomic profiles, interestingly these methods rely on a much less stringent type of annotated datasets compared to fully supervised learning methods. Recent years have witnessed an exponential increase in the amount of single-cell RNASeq data generated, particularly in studies of development. One of the major challenges is to identify individual cell types within the data. Expert knowledge is required to identify the relevant marker genes, tissue and timing that will enable the cell type identification. This information can be difficult to obtain and calls for automated cell type classification approaches. Classical classification techniques would solve this problem by training a classifier on a fully supervised dataset. However, this only pushes the problem further, as a dataset annotated at single-cell resolution is still needed for training. Here we propose instead to take advantage of the partial label learning framework which let us train our classifiers on a set of candidate labels per transcriptomic profile. This approach overcomes the need for a training dataset annotated at single-cell resolution. We show that we obtain classification accuracy similar to the fully supervised case. We explore the effect of varying the amount of partially labeled data and of considering the hierarchical structure of the label set (derived from the developmental processes) in the models on simulated and real biological datasets.
- Open Access
Reconstructing Voice Identity from Noninvasive Auditory Cortex Recordings
The cerebral processing of voice information is known to engage, in human as well as non-human primates, “temporal voice areas” (TVAs) that respond preferentially to conspecific vocalizations. However, how voice information is represented by neuronal populations in these areas, particularly speaker identity information, remains poorly understood. Here, we used a deep neural network (DNN) to generate a high-level, small-dimension representational space for voice identity—the ‘voice latent space’ (VLS)—and examined its linear relation with cerebral activity via encoding, representational similarity, and decoding analyses. We find that the VLS maps onto fMRI measures of cerebral activity in response to tens of thousands of voice stimuli from hundreds of different speaker identities and better accounts for the representational geometry for speaker identity in the TVAs than in A1. Moreover, the VLS allowed TVA-based reconstructions of voice stimuli that preserved essential aspects of speaker identity as assessed by both machine classifiers and human listeners. These results indicate that the DNN-derived VLS provides high-level representations of voice identity information in the TVAs.
Open AccessReconstructing Voice Identity from Noninvasive Auditory Cortex Recordings
The cerebral processing of voice information is known to engage, in human as well as non-human primates, “temporal voice areas” (TVAs) that respond preferentially to conspecific vocalizations. However, how voice information is represented by neuronal populations in these areas, particularly speaker identity information, remains poorly understood. Here, we used a deep neural network (DNN) to generate a high-level, small-dimension representational space for voice identity—the ‘voice latent space’ (VLS)—and examined its linear relation with cerebral activity via encoding, representational similarity, and decoding analyses. We find that the VLS maps onto fMRI measures of cerebral activity in response to tens of thousands of voice stimuli from hundreds of different speaker identities and better accounts for the representational geometry for speaker identity in the TVAs than in A1. Moreover, the VLS allowed TVA-based reconstructions of voice stimuli that preserved essential aspects of speaker identity as assessed by both machine classifiers and human listeners. These results indicate that the DNN-derived VLS provides high-level representations of voice identity information in the TVAs.
- Open Access
Site Agnostic Approach to Early Detection of Cyberbullying on Social Media Networks
The rise in the use of social media networks has increased the prevalence of cyberbullying, and time is paramount to reduce the negative effects that derive from those behaviours on any social media platform. This paper aims to study the early detection problem from a general perspective by carrying out experiments over two independent datasets (Instagram and Vine), exclusively using users’ comments. We used textual information from comments over baseline early detection models (fixed, threshold, and dual models) to apply three different methods of improving early detection. First, we evaluated the performance of Doc2Vec features. Finally, we also presented multiple instance learning (MIL) on early detection models and we assessed its performance. We applied () as an early detection metric to asses the performance of the presented methods. We conclude that the inclusion of Doc2Vec features improves the performance of baseline early detection models by up to 79.6%. Moreover, multiple instance learning shows an important positive effect for the Vine dataset, where smaller post sizes and less use of the English language are present, with a further improvement of up to 13%, but no significant enhancement is shown for the Instagram dataset.Open AccessSite Agnostic Approach to Early Detection of Cyberbullying on Social Media Networks
The rise in the use of social media networks has increased the prevalence of cyberbullying, and time is paramount to reduce the negative effects that derive from those behaviours on any social media platform. This paper aims to study the early detection problem from a general perspective by carrying out experiments over two independent datasets (Instagram and Vine), exclusively using users’ comments. We used textual information from comments over baseline early detection models (fixed, threshold, and dual models) to apply three different methods of improving early detection. First, we evaluated the performance of Doc2Vec features. Finally, we also presented multiple instance learning (MIL) on early detection models and we assessed its performance. We applied () as an early detection metric to asses the performance of the presented methods. We conclude that the inclusion of Doc2Vec features improves the performance of baseline early detection models by up to 79.6%. Moreover, multiple instance learning shows an important positive effect for the Vine dataset, where smaller post sizes and less use of the English language are present, with a further improvement of up to 13%, but no significant enhancement is shown for the Instagram dataset. - 1 December 2020
- Published by Elsevier in Computer Methods and Programs in Biomedicine
- Vol. 197, 105730
Inter-subject pattern analysis for multivariate group analysis of functional neuroimaging. A unifying formalization
- 1 December 2020
- Published by Elsevier in Computer Methods and Programs in Biomedicine
- Vol. 197, 105730
Inter-subject pattern analysis for multivariate group analysis of functional neuroimaging. A unifying formalization
- Open Access
Inter-subject pattern analysis for multivariate group analysis of functional neuroimaging. A unifying formalization
Background and objective: In medical imaging, population studies have to overcome the differences that exist between individuals to identify invariant image features that can be used for diagnosis purposes. In functional neuroimaging, an appealing solution to identify neural coding principles that hold at the population level is inter-subject pattern analysis, i.e. to learn a predictive model on data from multiple subjects and evaluate its generalization performance on new subjects. Although it has gained popularity in recent years, its widespread adoption is still hampered by the blatant lack of a formal definition in the literature. In this paper, we precisely introduce the first principled formalization of inter-subject pattern analysis targeted at multivariate group analysis of functional neuroimaging. Methods: We propose to frame inter-subject pattern analysis as a multi-source transductive transfer question, thus grounding it within several well defined machine learning settings and broadening the spectrum of usable algorithms. We describe two sets of inter-subject brain decoding experiments that use several open datasets: a magnetoencephalography study with 16 subjects and a functional magnetic resonance imaging paradigm with 100 subjects. We assess the relevance of our framework by performing model comparisons, where one brain decoding model exploits our formalization while others do not. Results: The first set of experiments demonstrates the superiority of a brain decoder that uses subject-by-subject standardization compared to state of the art models that use other standardization schemes, making the case for the interest of the transductive and the multi-source components of our formalization The second set of experiments quantitatively shows that, even after such transformation, it is more difficult for a brain decoder to generalize to new participants rather than to new data from participants available in the training phase, thus highlighting the transfer gap that needs to be overcome. Conclusion: This paper describes the first formalization of inter-subject pattern analysis as a multi-source transductive transfer learning problem. We demonstrate the added value of this formalization using proof-of-concept experiments on several complementary functional neuroimaging datasets. This work should contribute to popularize inter-subject pattern analysis for functional neuroimaging population studies and pave the road for future methodological innovations.
Open AccessInter-subject pattern analysis for multivariate group analysis of functional neuroimaging. A unifying formalization
Background and objective: In medical imaging, population studies have to overcome the differences that exist between individuals to identify invariant image features that can be used for diagnosis purposes. In functional neuroimaging, an appealing solution to identify neural coding principles that hold at the population level is inter-subject pattern analysis, i.e. to learn a predictive model on data from multiple subjects and evaluate its generalization performance on new subjects. Although it has gained popularity in recent years, its widespread adoption is still hampered by the blatant lack of a formal definition in the literature. In this paper, we precisely introduce the first principled formalization of inter-subject pattern analysis targeted at multivariate group analysis of functional neuroimaging. Methods: We propose to frame inter-subject pattern analysis as a multi-source transductive transfer question, thus grounding it within several well defined machine learning settings and broadening the spectrum of usable algorithms. We describe two sets of inter-subject brain decoding experiments that use several open datasets: a magnetoencephalography study with 16 subjects and a functional magnetic resonance imaging paradigm with 100 subjects. We assess the relevance of our framework by performing model comparisons, where one brain decoding model exploits our formalization while others do not. Results: The first set of experiments demonstrates the superiority of a brain decoder that uses subject-by-subject standardization compared to state of the art models that use other standardization schemes, making the case for the interest of the transductive and the multi-source components of our formalization The second set of experiments quantitatively shows that, even after such transformation, it is more difficult for a brain decoder to generalize to new participants rather than to new data from participants available in the training phase, thus highlighting the transfer gap that needs to be overcome. Conclusion: This paper describes the first formalization of inter-subject pattern analysis as a multi-source transductive transfer learning problem. We demonstrate the added value of this formalization using proof-of-concept experiments on several complementary functional neuroimaging datasets. This work should contribute to popularize inter-subject pattern analysis for functional neuroimaging population studies and pave the road for future methodological innovations.