DeepAdversaries: Examining the Robustness of Deep Learning Models for Galaxy Morphology Classification. Ćiprijanović, A., Kafkes, D., Snyder, G., Sánchez, F. J., Perdue, G. N., Pedro, K., Nord, B., Madireddy, S., & Wild, S. M. arXiv:2112.14299 [astro-ph], December, 2021. arXiv: 2112.14299![DeepAdversaries: Examining the Robustness of Deep Learning Models for Galaxy Morphology Classification [link]](https://bibbase.org/img/filetypes/link.svg "link")
Paper abstract bibtex Data processing and analysis pipelines in cosmological survey experiments introduce data perturbations that can significantly degrade the performance of deep learning-based models. Given the increased adoption of supervised deep learning methods for processing and analysis of cosmological survey data, the assessment of data perturbation effects and the development of methods that increase model robustness are increasingly important. In the context of morphological classification of galaxies, we study the effects of perturbations in imaging data. In particular, we examine the consequences of using neural networks when training on baseline data and testing on perturbed data. We consider perturbations associated with two primary sources: 1) increased observational noise as represented by higher levels of Poisson noise and 2) data processing noise incurred by steps such as image compression or telescope errors as represented by one-pixel adversarial attacks. We also test the efficacy of domain adaptation techniques in mitigating the perturbation-driven errors. We use classification accuracy, latent space visualizations, and latent space distance to assess model robustness. Without domain adaptation, we find that processing pixel-level errors easily flip the classification into an incorrect class and that higher observational noise makes the model trained on low-noise data unable to classify galaxy morphologies. On the other hand, we show that training with domain adaptation improves model robustness and mitigates the effects of these perturbations, improving the classification accuracy by 23% on data with higher observational noise. Domain adaptation also increases by a factor of \textasciitilde2.3 the latent space distance between the baseline and the incorrectly classified one-pixel perturbed image, making the model more robust to inadvertent perturbations.
@article{ciprijanovic_deepadversaries_2021,
title = {{DeepAdversaries}: {Examining} the {Robustness} of {Deep} {Learning} {Models} for {Galaxy} {Morphology} {Classification}},
shorttitle = {{DeepAdversaries}},
url = {http://arxiv.org/abs/2112.14299},
abstract = {Data processing and analysis pipelines in cosmological survey experiments introduce data perturbations that can significantly degrade the performance of deep learning-based models. Given the increased adoption of supervised deep learning methods for processing and analysis of cosmological survey data, the assessment of data perturbation effects and the development of methods that increase model robustness are increasingly important. In the context of morphological classification of galaxies, we study the effects of perturbations in imaging data. In particular, we examine the consequences of using neural networks when training on baseline data and testing on perturbed data. We consider perturbations associated with two primary sources: 1) increased observational noise as represented by higher levels of Poisson noise and 2) data processing noise incurred by steps such as image compression or telescope errors as represented by one-pixel adversarial attacks. We also test the efficacy of domain adaptation techniques in mitigating the perturbation-driven errors. We use classification accuracy, latent space visualizations, and latent space distance to assess model robustness. Without domain adaptation, we find that processing pixel-level errors easily flip the classification into an incorrect class and that higher observational noise makes the model trained on low-noise data unable to classify galaxy morphologies. On the other hand, we show that training with domain adaptation improves model robustness and mitigates the effects of these perturbations, improving the classification accuracy by 23\% on data with higher observational noise. Domain adaptation also increases by a factor of {\textasciitilde}2.3 the latent space distance between the baseline and the incorrectly classified one-pixel perturbed image, making the model more robust to inadvertent perturbations.},
urldate = {2022-01-05},
journal = {arXiv:2112.14299 [astro-ph]},
author = {Ćiprijanović, Aleksandra and Kafkes, Diana and Snyder, Gregory and Sánchez, F. Javier and Perdue, Gabriel Nathan and Pedro, Kevin and Nord, Brian and Madireddy, Sandeep and Wild, Stefan M.},
month = dec,
year = {2021},
note = {arXiv: 2112.14299},
keywords = {Astrophysics - Astrophysics of Galaxies, Computer Science - Artificial Intelligence, Computer Science - Computer Vision and Pattern Recognition, Computer Science - Machine Learning},
}
Downloads: 0
{"_id":"aJzx5eZMAopyLpdDw","bibbaseid":"iprijanovi-kafkes-snyder-snchez-perdue-pedro-nord-madireddy-etal-deepadversariesexaminingtherobustnessofdeeplearningmodelsforgalaxymorphologyclassification-2021","author_short":["Ćiprijanović, A.","Kafkes, D.","Snyder, G.","Sánchez, F. J.","Perdue, G. N.","Pedro, K.","Nord, B.","Madireddy, S.","Wild, S. M."],"bibdata":{"bibtype":"article","type":"article","title":"DeepAdversaries: Examining the Robustness of Deep Learning Models for Galaxy Morphology Classification","shorttitle":"DeepAdversaries","url":"http://arxiv.org/abs/2112.14299","abstract":"Data processing and analysis pipelines in cosmological survey experiments introduce data perturbations that can significantly degrade the performance of deep learning-based models. Given the increased adoption of supervised deep learning methods for processing and analysis of cosmological survey data, the assessment of data perturbation effects and the development of methods that increase model robustness are increasingly important. In the context of morphological classification of galaxies, we study the effects of perturbations in imaging data. In particular, we examine the consequences of using neural networks when training on baseline data and testing on perturbed data. We consider perturbations associated with two primary sources: 1) increased observational noise as represented by higher levels of Poisson noise and 2) data processing noise incurred by steps such as image compression or telescope errors as represented by one-pixel adversarial attacks. We also test the efficacy of domain adaptation techniques in mitigating the perturbation-driven errors. We use classification accuracy, latent space visualizations, and latent space distance to assess model robustness. Without domain adaptation, we find that processing pixel-level errors easily flip the classification into an incorrect class and that higher observational noise makes the model trained on low-noise data unable to classify galaxy morphologies. On the other hand, we show that training with domain adaptation improves model robustness and mitigates the effects of these perturbations, improving the classification accuracy by 23% on data with higher observational noise. Domain adaptation also increases by a factor of \\textasciitilde2.3 the latent space distance between the baseline and the incorrectly classified one-pixel perturbed image, making the model more robust to inadvertent perturbations.","urldate":"2022-01-05","journal":"arXiv:2112.14299 [astro-ph]","author":[{"propositions":[],"lastnames":["Ćiprijanović"],"firstnames":["Aleksandra"],"suffixes":[]},{"propositions":[],"lastnames":["Kafkes"],"firstnames":["Diana"],"suffixes":[]},{"propositions":[],"lastnames":["Snyder"],"firstnames":["Gregory"],"suffixes":[]},{"propositions":[],"lastnames":["Sánchez"],"firstnames":["F.","Javier"],"suffixes":[]},{"propositions":[],"lastnames":["Perdue"],"firstnames":["Gabriel","Nathan"],"suffixes":[]},{"propositions":[],"lastnames":["Pedro"],"firstnames":["Kevin"],"suffixes":[]},{"propositions":[],"lastnames":["Nord"],"firstnames":["Brian"],"suffixes":[]},{"propositions":[],"lastnames":["Madireddy"],"firstnames":["Sandeep"],"suffixes":[]},{"propositions":[],"lastnames":["Wild"],"firstnames":["Stefan","M."],"suffixes":[]}],"month":"December","year":"2021","note":"arXiv: 2112.14299","keywords":"Astrophysics - Astrophysics of Galaxies, Computer Science - Artificial Intelligence, Computer Science - Computer Vision and Pattern Recognition, Computer Science - Machine Learning","bibtex":"@article{ciprijanovic_deepadversaries_2021,\n\ttitle = {{DeepAdversaries}: {Examining} the {Robustness} of {Deep} {Learning} {Models} for {Galaxy} {Morphology} {Classification}},\n\tshorttitle = {{DeepAdversaries}},\n\turl = {http://arxiv.org/abs/2112.14299},\n\tabstract = {Data processing and analysis pipelines in cosmological survey experiments introduce data perturbations that can significantly degrade the performance of deep learning-based models. Given the increased adoption of supervised deep learning methods for processing and analysis of cosmological survey data, the assessment of data perturbation effects and the development of methods that increase model robustness are increasingly important. In the context of morphological classification of galaxies, we study the effects of perturbations in imaging data. In particular, we examine the consequences of using neural networks when training on baseline data and testing on perturbed data. We consider perturbations associated with two primary sources: 1) increased observational noise as represented by higher levels of Poisson noise and 2) data processing noise incurred by steps such as image compression or telescope errors as represented by one-pixel adversarial attacks. We also test the efficacy of domain adaptation techniques in mitigating the perturbation-driven errors. We use classification accuracy, latent space visualizations, and latent space distance to assess model robustness. Without domain adaptation, we find that processing pixel-level errors easily flip the classification into an incorrect class and that higher observational noise makes the model trained on low-noise data unable to classify galaxy morphologies. On the other hand, we show that training with domain adaptation improves model robustness and mitigates the effects of these perturbations, improving the classification accuracy by 23\\% on data with higher observational noise. Domain adaptation also increases by a factor of {\\textasciitilde}2.3 the latent space distance between the baseline and the incorrectly classified one-pixel perturbed image, making the model more robust to inadvertent perturbations.},\n\turldate = {2022-01-05},\n\tjournal = {arXiv:2112.14299 [astro-ph]},\n\tauthor = {Ćiprijanović, Aleksandra and Kafkes, Diana and Snyder, Gregory and Sánchez, F. Javier and Perdue, Gabriel Nathan and Pedro, Kevin and Nord, Brian and Madireddy, Sandeep and Wild, Stefan M.},\n\tmonth = dec,\n\tyear = {2021},\n\tnote = {arXiv: 2112.14299},\n\tkeywords = {Astrophysics - Astrophysics of Galaxies, Computer Science - Artificial Intelligence, Computer Science - Computer Vision and Pattern Recognition, Computer Science - Machine Learning},\n}\n\n","author_short":["Ćiprijanović, A.","Kafkes, D.","Snyder, G.","Sánchez, F. J.","Perdue, G. N.","Pedro, K.","Nord, B.","Madireddy, S.","Wild, S. M."],"key":"ciprijanovic_deepadversaries_2021","id":"ciprijanovic_deepadversaries_2021","bibbaseid":"iprijanovi-kafkes-snyder-snchez-perdue-pedro-nord-madireddy-etal-deepadversariesexaminingtherobustnessofdeeplearningmodelsforgalaxymorphologyclassification-2021","role":"author","urls":{"Paper":"http://arxiv.org/abs/2112.14299"},"keyword":["Astrophysics - Astrophysics of Galaxies","Computer Science - Artificial Intelligence","Computer Science - Computer Vision and Pattern Recognition","Computer Science - Machine Learning"],"metadata":{"authorlinks":{}}},"bibtype":"article","biburl":"https://bibbase.org/zotero/polyphant","dataSources":["7gvjSdWrEu7z5vjjj"],"keywords":["astrophysics - astrophysics of galaxies","computer science - artificial intelligence","computer science - computer vision and pattern recognition","computer science - machine learning"],"search_terms":["deepadversaries","examining","robustness","deep","learning","models","galaxy","morphology","classification","ćiprijanović","kafkes","snyder","sánchez","perdue","pedro","nord","madireddy","wild"],"title":"DeepAdversaries: Examining the Robustness of Deep Learning Models for Galaxy Morphology Classification","year":2021}