MDMariia Drozdova

Mariia Drozdova

*This is an algorithmically generated author record based on publicly available information.

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article


Mariia Drozdova has a total of 28 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Open Access

    Can Continuous-Time Diffusion Models Generate and Solve Globally Constrained Discrete Problems? A Study on Sudoku

    by Mariia Drozdova

    Can standard continuous-time generative models represent distributions whose support is an extremely sparse, globally constrained discrete set? We study this question using completed Sudoku grids as a controlled testbed, treating them as a subset of a continuous relaxation space. We train flow-matching and score-based models along a Gaussian probability path and compare deterministic (ODE) sampling, stochastic (SDE) sampling, and DDPM-style discretizations derived from the same continuous-time training. Unconditionally, stochastic sampling substantially outperforms deterministic flows; score-based samplers are the most reliable among continuous-time methods, and DDPM-style ancestral sampling achieves the highest validity overall. We further show that the same models can be repurposed for guided generation: by repeatedly sampling completions under clamped clues and stopping when constraints are satisfied, the model acts as a probabilistic Sudoku solver. Although far less sample-efficient than classical solvers and discrete-geometry-aware diffusion methods, these experiments demonstrate that classic diffusion/flow formulations can assign non-zero probability mass to globally constrained combinatorial structures and can be used for constraint satisfaction via stochastic search.

  • Open Access

    Examining vision foundation models for classification and detection in optical and radio astronomy

    by E. Lastufka, O. Bait, M. Drozdova, V. Kinakh, D. Piras, M. Audard, M. Dessauges-Zavadsky, T. Holotyak, D. Schaerer, S. Voloshynovskiy

    Context . Vision foundation models, which have demonstrated significant potential in many multimedia applications, are often underutilized in the natural sciences. This is primarily due to mismatches between the nature of domain-specific scientific data and the typical training data used for foundation models, leading to distribution shifts. Scientific data often differ substantially in structure and characteristics, and researchers frequently face the challenge of optimizing model performance with limited labeled data of only a few hundred or thousand images. Aims . This work evaluates the performance of vision foundation models in astrophysics, with a focus on identifying the best practices for adapting these models to domain-specific datasets. We aim to establish a framework for selecting, fine-tuning, and optimizing these models for common tasks in optical and radio astronomy. Methods . We compared multiple foundation models, including self-supervised, weakly supervised, and distillation-based architectures, across two representative optical and radio datasets. Experiments involved different fine-tuning strategies, projector heads, and data preprocessing techniques, with performance evaluated on classification and detection metrics. Results . Features extracted by specific foundation models improved classification accuracy for optical galaxy images compared to conventional supervised training. Similarly, these models achieved equivalent or superior performance in object detection tasks with radio images. However, classification performance for radio galaxy images was generally poor, often falling short of traditional supervised approaches. Conclusions . These findings suggest that selecting suitable vision foundation models for astrophysics applications requires careful consideration of the model characteristics and alignment with the specific requirements of the downstream tasks. This study demonstrates that vision foundation models can be effectively adapted to astrophysical applications, provided practitioners iterate on model selection, training strategies, and data handling. The proposed framework bridges the gap between these advanced models and the unique demands of astronomy, enabling broader adoption of deep learning in the field.


  • Open Access

    Visualizing three years of STIX X-ray flare observations using self-supervised learning

    by Mariia Drozdova, Vitaliy Kinakh, Francesco Ramunno, Erica Lastufka, Slava Voloshynovskiy

    Operating continuously for over three years, Solar Orbiter's STIX has observed more than 43 thousand X-ray flares. This study presents a compelling visualization of this publicly available database, using self-supervised learning to organize reconstructed flare images by their visual properties. Networks designed for self-supervised learning, such as Masked Siamese Networks or Autoencoders, are able to learn latent space embeddings which encode core characteristics of the data. We investigate the effectiveness of various pre-trained vision models, fine-tuning strategies, and image preparation. This visual representation offers a valuable starting point for identifying interesting events and grouping flares based on shared morphological characteristics, useful for conducting statistical studies or finding unique flares in this rich set of observations.

  • Open Access

    Self-supervised learning on MeerKAT wide-field continuum images

    by E. Lastufka, O. Bait, O. Taran, M. Drozdova, V. Kinakh, D. Piras, M. Audard, M. Dessauges-Zavadsky, T. Holotyak, D. Schaerer, et al.

    Context. Self-supervised learning (SSL) applied to natural images has demonstrated a remarkable ability to learn meaningful, low-dimension representations without labels, resulting in models that are adaptable to many different tasks. Until now, applications of SSL to astronomical images have been limited to Galaxy Zoo datasets, which require a significant amount of preprocessing to prepare sparse images centered on a single galaxy. With wide-field survey instruments at the forefront of the Square Kilometer Array (SKA) era, this approach to gathering training data is impractical. Aims. We demonstrate that continuum images from surveys such as the MeerKAT Galactic Cluster Legacy Survey (MGCLS) can be successfully used with SSL, without extracting single-galaxy cutouts. Methods. Using the SSL framework DINO, we experimented with various preprocessing steps, augmentations, and architectures to determine the optimal approach for this data. We trained both ResNet50 and Vision Transformer (ViT) backbones. Results. Our models match state-of-the-art results (trained on Radio Galaxy Zoo) for FRI/FRII morphology classification. Furthermore, they predict the number of compact sources via linear regression with much higher accuracy. Open-source foundation models trained on natural images such as DINOv2 also excel at simple FRI/FRII classification; the advantage of domain-specific backbones is much smaller models trained on far less data. Smaller models are more efficient to fine-tune, and doing so results in a similar performance between our models, the state-of-the-art, and open-source models on multi-class morphology classification. Conclusions. Using source-rich crops from wide-field images to train multi-purpose models is an easily scalable approach that significantly reduces data preparation time. For the tasks evaluated in this work, twenty thousand crops is sufficient training data for models that produce results similar to state-of-the-art. In the future, complex tasks like source detection and characterization, together with domain-specific tasks, ought to demonstrate the true advantages of training models with radio astronomy data over natural-image foundation models.


  • Open Access

    Solar synthetic imaging: Introducing denoising diffusion probabilistic models on SDO/AIA data

    by F. P. Ramunno, S. Hackstein, V. Kinakh, M. Drozdova, G. Quétant, A. Csillaghy, S. Voloshynovskiy

    For the luck of humanity, there are way less big solar flares than small ones. Even if these are good news, this makes it challenging to train machine learning algorithms able to model solar activity. As a result, solar monitoring applications, including flare forecasting, suffer from this lack of input data. To overcome this issue, generative deep learning models can be utilised to produce synthetic images representing solar activity and thus compensating the rarity of big events. This study aims to develop a method that can generate synthetic images of the Sun with the ability to include flare of a specific intensity. To achieve our goals, we introduce a Denoising Diffusion Probabilistic Model (DDPM). We train it with a carefully crafted dataset from the Atmospheric Image Assembly (AIA) instrument on the SDO spacecraft, specifically the 171 Å band, which captures images of coronal loops, filaments, flares, and active regions. GOES X-ray measurements are employed to classify each image based on the solar flare scale (A, B, C, M, X), after selecting the flaring images from AIA using the Heliophysics Event Knowledgebase, which allows for temporal localisation of the flaring events. The generative model performance is evaluated using cluster metrics, Fréchet Inception Distance (FID), and the F1-score. We demonstrate state-of-the-art results in generating solar images and conduct two experiments that use the synthetic images. The first experiment trains a supervised classifier to identify those events. The second experiment trains a basic solar flare predictor. The experiments demonstrate the effectiveness of additional synthetic samples to addressing the problem of imbalanced datasets. We believe this is only the beginning of DDPM use with solar data. It remains to gain a better understanding of the generation capabilities of the denoising diffusion probabilistic models in the contest of solar flare predictions and apply them to other deep learning and physical tasks, such as AIA to HMI () image translation.


  • Open Access

    MV–MR: Multi-Views and Multi-Representations for Self-Supervised Learning and Knowledge Distillation

    by Vitaliy Kinakh, Mariia Drozdova, Slava Voloshynovskiy

    We present a new method of self-supervised learning and knowledge distillation based on multi-views and multi-representations (MV–MR). MV–MR is based on the maximization of dependence between learnable embeddings from augmented and non-augmented views, jointly with the maximization of dependence between learnable embeddings from the augmented view and multiple non-learnable representations from the non-augmented view. We show that the proposed method can be used for efficient self-supervised classification and model-agnostic knowledge distillation. Unlike other self-supervised techniques, our approach does not use any contrastive learning, clustering, or stop gradients. MV–MR is a generic framework allowing the incorporation of constraints on the learnable embeddings via the usage of image multi-representations as regularizers. The proposed method is used for knowledge distillation. MV–MR provides state-of-the-art self-supervised performance on the STL10 and CIFAR20 datasets in a linear evaluation setup. We show that a low-complexity ResNet50 model pretrained using proposed knowledge distillation based on the CLIP ViT model achieves state-of-the-art performance on STL10 and CIFAR100 datasets.


  • Open Access

    Radio-astronomical image reconstruction with a conditional denoising diffusion model

    by M. Drozdova, V. Kinakh, O. Bait, O. Taran, E. Lastufka, M. Dessauges-Zavadsky, T. Holotyak, D. Schaerer, S. Voloshynovskiy

    Context.Reconstructing sky models from dirty radio images for accurate source extraction, including source localization and flux estimation, is a complex yet critical task, and has important applications in galaxy evolution studies at high redshift, particularly in deep extragalactic fields using for example the Atacama Large Millimetre Array (ALMA). With the development of large-scale projects, such as the Square Kilometre Array (SKA), we anticipate the need for more advanced source-extraction methods. Existing techniques, such as CLEAN and PyBDSF, currently struggle to effectively extract faint sources, highlighting the necessity for the development of more precise and robust methods.Aims.The success of the source-extraction process critically depends on the quality and accuracy of image reconstruction. As the imaging process represents an “information-lossy” operator, the reconstruction is characterized by uncertainty. The current study proposes the application of stochastic neural networks for the direct reconstruction of sky models from “dirty” images. This approach allows us to localize radio sources and to determine their fluxes with corresponding uncertainties, providing a potential advancement in the field of radio-source characterization.Methods.We used a dataset of 10 164 images simulated with the CASA tool based on the ALMA Cycle 5.3 antenna configuration. We applied conditional denoising diffusion probabilistic models (DDPMs) to directly reconstruct sky models from these dirty images, and then processed these models using Photutils to extract the coordinates and fluxes of the sources. To test the robustness of the proposed model, which was trained on a fixed water vapor value, we examined its performance under varying levels of water vapor.Results.We demonstrate that the proposed approach is state of the art in terms of source localisation, achieving over 90% completeness at a signal-to-noise ratio (S/N) of as low as 2. Additionally, the described method offers an inherent measure of prediction reliability thanks to the stochastic nature of the chosen model. In terms of flux estimation, the proposed model surpasses PyBDSF in terms of performance, accurately extracting fluxes for 96% of the sources in the test set, a notable improvement over the 57% achieved by CLEAN+ PyBDSF.Conclusions.Conditional DDPMs are a powerful tool for image-to-image translation, yielding accurate and robust characterization of radio sources, and outperforming existing methodologies. While this study underscores the significant potential of DDPMs for applications in radio astronomy, we also acknowledge certain limitations that accompany their use, and suggest directions for further refinement and research.


  • Open Access

    Radio-astronomical Image Reconstruction with Conditional Denoising Diffusion Model

    by Mariia Drozdova, Vitaliy Kinakh, Omkar Bait, Olga Taran, Erica Lastufka, Miroslava Dessauges-Zavadsky, Taras Holotyak, Daniel Schaerer, Slava Voloshynovskiy

    Reconstructing sky models from dirty radio images for accurate source localization and flux estimation is crucial for studying galaxy evolution at high redshift, especially in deep fields using instruments like the Atacama Large Millimetre Array (ALMA). With new projects like the Square Kilometre Array (SKA), there's a growing need for better source extraction methods. Current techniques, such as CLEAN and PyBDSF, often fail to detect faint sources, highlighting the need for more accurate methods. This study proposes using stochastic neural networks to rebuild sky models directly from dirty images. This method can pinpoint radio sources and measure their fluxes with related uncertainties, marking a potential improvement in radio source characterization. We tested this approach on 10164 images simulated with the CASA tool simalma, based on ALMA's Cycle 5.3 antenna setup. We applied conditional Denoising Diffusion Probabilistic Models (DDPMs) for sky models reconstruction, then used Photutils to determine source coordinates and fluxes, assessing the model's performance across different water vapor levels. Our method showed excellence in source localization, achieving more than 90% completeness at a signal-to-noise ratio (SNR) as low as 2. It also surpassed PyBDSF in flux estimation, accurately identifying fluxes for 96% of sources in the test set, a significant improvement over CLEAN+ PyBDSF's 57%. Conditional DDPMs is a powerful tool for image-to-image translation, yielding accurate and robust characterisation of radio sources, and outperforming existing methodologies. While this study underscores its significant potential for applications in radio astronomy, we also acknowledge certain limitations that accompany its usage, suggesting directions for further refinement and research.

  • Open Access

    Hubble Meets Webb: Image-to-Image Translation in Astronomy

    by Vitaliy Kinakh, Yury Belousov, Guillaume Quétant, Mariia Drozdova, Taras Holotyak, Daniel Schaerer, Slava Voloshynovskiy

    This work explores the generation of James Webb Space Telescope (JWSP) imagery via image-to-image translation from the available Hubble Space Telescope (HST) data. Comparative analysis encompasses the Pix2Pix, CycleGAN, TURBO, and DDPM-based Palette methodologies, assessing the criticality of image registration in astronomy. While the focus of this study is not on the scientific evaluation of model fairness, we note that the techniques employed may bear some limitations and the translated images could include elements that are not present in actual astronomical phenomena. To mitigate this, uncertainty estimation is integrated into our methodology, enhancing the translation’s integrity and assisting astronomers in distinguishing between reliable predictions and those of questionable certainty. The evaluation was performed using metrics including MSE, SSIM, PSNR, LPIPS, and FID. The paper introduces a novel approach to quantifying uncertainty within image translation, leveraging the stochastic nature of DDPMs. This innovation not only bolsters our confidence in the translated images but also provides a valuable tool for future astronomical experiment planning. By offering predictive insights when JWST data are unavailable, our approach allows for informed preparatory strategies for making observations with the upcoming JWST, potentially optimizing its precious observational resources. To the best of our knowledge, this work is the first attempt to apply image-to-image translation for astronomical sensor-to-sensor translation.