Publications

@misc{semenkov2026interpretablemegdecodingperceived,
      title={Interpretable MEG Decoding of Perceived Speech: Cortical Sources and the Stimulus Features That Drive Retrieval}, 
      author={Ilia Semenkov and Daria Kleeva and Ivan Dakhtin and Zarina Maksudova and Alex Ossadtchi},
      year={2026},
      eprint={2608.01481},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      doi={10.48550/arXiv.2608.01481},
      url={https://arxiv.org/abs/2608.01481}, 
}

@inproceedings{10.1145/3812835.3814867,
author = {Semenkov, Ilia and Nikiforova, Sophia and Fedosov, Nikita and Sazonov, Oleg and Medvedeva, Daria and Nekrasova, Uliya and Ossadtchi, Alex},
title = {Poster: Geometry-Aware Real-Time Finger Decoding from Residual-Limb sEMG},
year = {2026},
isbn = {9798400727115},
publisher = {Association for Computing Machinery},
address = {New York, NY, USA},
url = {https://doi.org/10.1145/3812835.3814867},
doi = {10.1145/3812835.3814867},
abstract = {Interactive upper-limb prostheses require continuous decoding of rich finger kinematics under tight latency, memory, and model-size constraints. We study prediction of 20 finger joint angles from six-channel residual-limb sEMG collected from five transradial amputees during VR-based pre-prosthetic training. We compare two front ends under the same MLP regressor: tangent-space embeddings of windowed EMG covariance matrices and a CNN feature extractor. In blocked cross-validation, the tangent-space front end outperforms the CNN baseline (Pearson 0.80 vs. 0.71; lower MAE and circular standard deviation) while reducing trainable parameters by 34\% (177k vs. 268k) and memory use from 1.02 MB to 0.68 MB. Per-window processing is sub-millisecond, leaving headroom for interactive-rate operation. These results support geometry-aware covariance features as a compact front end for myoelectric interfaces.},
booktitle = {Proceedings of the 24th Annual International Conference on Mobile Systems, Applications and Services Companion},
pages = {35–36},
numpages = {2},
keywords = {wearable computing, electromyography, myoelectric control, realtime inference, embedded machine learning},
location = {University of Cambridge, Cambridge, United Kingdom},
series = {MobiSys Companion '26}
}

@INPROCEEDINGS{11575012,
  author={Maksudova, Zarina and Shestakov, Denis and Semenkov, Ilia and Bernovek, Anastasia and Makarov, Ilya and Ossadtchi, Alex},
  booktitle={2026 ACM/IEEE International Conference on Embedded Artificial Intelligence and Sensing Systems - Posters and Demos (SenSys-Adjunct)}, 
  title={Poster Abstract: Physiology-Aware Foundation Models for Robust and Interpretable EEG Sensing}, 
  year={2026},
  volume={},
  number={},
  pages={83-84},
  keywords={Modeling;Electroencephalography;Standards;Learning (artificial intelligence);Printing;Filtering;Foundation models;Visualization;EEG;Interpretability;Foundation Models;SelfSupervised Learning;Sensing Systems},
  doi={10.1109/SenSys-Adjunct71932.2026.00051}}

@article{Komissarenko2025,
doi = {10.1038/s41597-025-05966-x},
url = {https://doi.org/10.1038/s41597-025-05966-x},
year = {2025},
volume = {12},
number = {1},
pages = {1687},
author = {Komissarenko, Anna and Voloshina, Ekaterina and Cheveleva, Anastasia and Semenkov, Ilia and Serikov, Oleg and Ossadtchi, Alex},
title = {SIGNAL: Dataset for Semantic and Inferred Grammar Neurological Analysis of Language},
journal = {Scientific Data},
abstract = {Recently, the idea of brain-model alignment has been the topic of several influential works. However, most of previous studies were based on datasets collected during regular reading tasks where the subjects were not exposed to processing linguistic incongruencies, and stimuli were not controlled for key linguistic properties. Meanwhile, interpretability studies of Large Language Models pay growing attention to thoroughly designed linguistic tasks based on certain acceptability measures. We present a dataset that contains 600 sentences with a combination of congruent and grammatically or/and semantically incongruent sentences coupled with high density 64-channel EEG recordings of 21 participants. The text stimuli were assessed by native speakers and later used in EEG recording and validation and LLM probing. The validation results proved suitability of the data for future research on brain-model alignment in the linguistic context.}
}

@article{Voskoboynikov_2025,
doi = {10.1088/1741-2552/adfc9c},
url = {https://doi.org/10.1088/1741-2552/adfc9c},
year = {2025},
month = {sep},
publisher = {IOP Publishing},
volume = {22},
number = {5},
pages = {056002},
author = {Voskoboynikov, Alexei and Aliverdiev, Magomed and Nekrasova, Yulia and Semenkov, Ilia and Skalnaya, Anastasia and Sinkin, Mikhail and Ossadtchi, Alexei},
title = {Towards stimulation-free automatic electrocorticographic speech mapping in neurosurgery patients},
journal = {Journal of Neural Engineering},
abstract = {Objective. The precise mapping of speech-related functions is crucial for successful neurosurgical interventions in epilepsy and brain tumor cases. Traditional methods like electrocortical stimulation mapping (ESM) are effective but carry a significant risk of inducing seizures. Methods. To address this, we have prepared a comprehensive ESM + electrocorticographic mapping (ECM) dataset from 14 patients with chronically implanted stereo-EEG electrodes. Then we explored several compact machine learning (ML) approaches to convert the ECM signals to the ground truth derived from the risky ESM procedure. Both procedures involved the standard picture naming task. As features, we used gamma-band power within successive temporal windows in the data averaged with respect to picture and voice onsets. We focused on a range of classifiers, including XGBoost, linear support vector classification (SVC), regularized logistic regression, random forest, k-nearest neighbors, decision tree, multi-Layer perceptron, AdaBoost and Gaussian Naive Bayes classifiers and equipped them with confidence interval estimates, crucial in a real-life application. We validated the ML approaches using a leave-one-patient-out procedure and computed ROC and Precision–Recall curves for various feature combinations. Results. For linear SVC we achieved ROC-AUC and PR-AUC scores of 0.91 and 0.88, respectively, which effectively distinguishes speech-related from non-related iEEG channels. We have also observed that the use of information on the voice onset moment notably improved the classification accuracy. Significance. We have for the first time rigorously compared the ECM and ESM results and mimicked a real-life use of the ECM technology. We have also provided public access to the comprehensive ECM+ESM dataset to pave the road towards safer and more reliable eloquent cortex mapping procedures.}
}

@article{OSSADTCHI2024120868,
title = {Representational dissimilarity component analysis (ReDisCA)},
journal = {NeuroImage},
volume = {301},
pages = {120868},
year = {2024},
issn = {1053-8119},
doi = {https://doi.org/10.1016/j.neuroimage.2024.120868},
url = {https://doi.org/10.1016/j.neuroimage.2024.120868},
author = {Alexei Ossadtchi and Ilia Semenkov and Anna Zhuravleva and Vladimir Kozunov and Oleg Serikov and Ekaterina Voloshina},
keywords = {EEG and MEG, Spatial–temporal decomposition, Representational similarity analysis, Source localization},
abstract = {The principle of Representational Similarity Analysis (RSA) posits that neural representations reflect the structure of encoded information, allowing exploration of spatial and temporal organization of brain information processing. Traditional RSA when applied to EEG or MEG data faces challenges in accessing activation time series at the brain source level due to modeling complexities and insufficient geometric/anatomical data. To overcome this, we introduce Representational Dissimilarity Component Analysis (ReDisCA), a method for estimating spatial–temporal components in EEG or MEG responses aligned with a target representational dissimilarity matrix (RDM). ReDisCA yields informative spatial filters and associated topographies, offering insights into the location of ”representationally relevant” sources. Applied to evoked response time series, ReDisCA produces temporal source activation profiles with the desired RDM. Importantly, while ReDisCA does not require inverse modeling its output is consistent with EEG and MEG observation equation and can be used as an input to rigorous source localization procedures. Demonstrating ReDisCA’s efficacy through simulations and comparison with conventional methods, we show superior source localization accuracy and apply the method to real EEG and MEG datasets, revealing physiologically plausible representational structures without inverse modeling. ReDisCA adds to the family of inverse modeling free methods such as independent component analysis (Makeig, 1995), Spatial spectral decomposition (Nikulin, 2011), and Source power comodulation (Dähne, 2014) designed for extraction sources with desired properties from EEG or MEG data. Extending its utility beyond EEG and MEG analysis, ReDisCA is likely to find application in fMRI data analysis and exploration of representational structures emerging in multilayered artificial neural networks.}
}

@ARTICLE{10380584,
  author={Semenkov, Ilia and Karpov, Aleksei and Savchenko, Andrey V. and Makarov, Ilya},
  journal={IEEE Access}, 
  title={Inpainting Semantic and Depth Features to Improve Visual Place Recognition in the Wild}, 
  year={2024},
  volume={12},
  number={},
  pages={5163-5176},
  keywords={Visualization;Semantics;Indexes;Image recognition;Refining;Pipelines;Vehicle dynamics;Image retrieval;Visual place recognition;image retrieval;inpainting;semantic segmentation;monocular depth estimation;NetVLAD;CosPlace},
  doi={10.1109/ACCESS.2024.3350038}}

@InProceedings{Dagaev_2024,
    author    = {Dagaev, Nikolay and Semenkov, Ilia and Ossadtchi, Alexei},
    title     = {EEG-Based fMRI Digital Twin: Towards a Cheap and Ecological Approach to Measure Subcortical Brain Activity},
    booktitle = {Proceedings of the 27th European Conference on Artificial Intelligence},
    month     = {October},
    year      = {2024},
    pages     = {4463-4466},
    doi       = {10.3233/FAIA241030},
}

@article{Semenkov_2023,
doi = {10.1088/1741-2552/acf7f3},
url = {https://doi.org/10.1088/1741-2552/acf7f3},
year = {2023},
month = {sep},
publisher = {IOP Publishing},
volume = {20},
number = {5},
pages = {056008},
author = {Semenkov, Ilia and Fedosov, Nikita and Makarov, Ilya and Ossadtchi, Alexei},
title = {Real-time low latency estimation of brain rhythms with deep neural networks},
journal = {Journal of Neural Engineering},
abstract = {Objective. Neurofeedback and brain-computer interfacing technology open the exciting opportunity for establishing interactive closed-loop real-time communication with the human brain. This requires interpreting brain’s rhythmic activity and generating timely feedback to the brain. Lower delay between neuronal events and the appropriate feedback increases the efficacy of such interaction. Novel more efficient approaches capable of tracking brain rhythm’s phase and envelope are needed for scenarios that entail instantaneous interaction with the brain circuits. Approach. Isolating narrow-band signals incurs fundamental delays. To some extent they can be compensated using forecasting models. Given the high quality of modern time series forecasting neural networks we explored their utility for low-latency extraction of brain rhythm parameters. We tested five neural networks with conceptually distinct architectures in forecasting synthetic EEG rhythms. The strongest architecture was then trained to simultaneously filter and forecast EEG data. We compared it against the state-of-the-art techniques using synthetic and real data from 25 subjects. Main results. The temporal convolutional network (TCN) remained the strongest forecasting model that achieved in the majority of testing scenarios 90% rhythm’s envelope correlation with 10 ms effective delay and  circular standard deviation of phase estimates. It also remained stable enough to noise level perturbations. Trained to filter and predict the TCN outperformed the cFIR, the Kalman filter based state-space estimation technique and remained on par with the larger Conv-TasNet architecture. Significance. Here we have for the first time demonstrated the utility of the neural network approach for low-latency narrow-band filtering of brain activity signals. Our proposed approach coupled with efficient implementation enhances the effectiveness of brain-state dependent paradigms across various applications. Moreover, our framework for forecasting EEG signals holds promise for investigating the predictability of brain activity, providing valuable insights into the fundamental questions surrounding the functional organization and hierarchical information processing properties of the brain.}
}

@article{10.1371/journal.pone.0275814,
    doi = {10.1371/journal.pone.0275814},
    author = {Dudyrev, Egor and Semenkov, Ilia and Kuznetsov, Sergei O. and Gusev, Gleb and Sharp, Andrew and Pianykh, Oleg S.},
    journal = {PLOS ONE},
    publisher = {Public Library of Science},
    title = {Human knowledge models: Learning applied knowledge from the data},
    year = {2022},
    month = {10},
    volume = {17},
    url = {https://doi.org/10.1371/journal.pone.0275814},
    pages = {1-16},
    abstract = {Artificial intelligence and machine learning have demonstrated remarkable results in science and applied work. However, present AI models, developed to be run on computers but used in human-driven applications, create a visible disconnect between AI forms of processing and human ways of discovering and using knowledge. In this work, we introduce a new concept of “Human Knowledge Models” (HKMs), designed to reproduce human computational abilities. Departing from a vast body of cognitive research, we formalized the definition of HKMs into a new form of machine learning. Then, by training the models with human processing capabilities, we learned human-like knowledge, that humans can not only understand, but also compute, modify, and apply. We used several datasets from different applied fields to demonstrate the advantages of HKMs, including their high predictive power and resistance to noise and overfitting. Our results proved that HKMs can efficiently mine knowledge directly from the data and can compete with complex AI models in explaining the main data patterns. As a result, our study reveals the great potential of HKMs, particularly in the decision-making applications where “black box” models cannot be accepted. Moreover, this improves our understanding of how well human decision-making, modeled by HKMs, can approach the ideal solutions in real-life problems.},
    number = {10},

}

@INPROCEEDINGS{9856964,
  author={Ershov, Egor and Savchik, Alex and Shepelev, Denis and Banić, Nikola and Brown, Michael S. and Timofte, Radu and Koščević, Karlo and Freeman, Michael and Tesalin, Vasily and Bocharov, Dmitry and Semenkov, Illya and Subašić, Marko and Lončarić, Sven and Terekhin, Arseniy and Liu, Shuai and Feng, Chaoyu and Wang, Hao and Zhu, Ran and Li, Yongqiang and Lei, Lei and Li, Zhihao and Yi, Si and Han, Ling-Hao and Wu, Ruiqi and Jin, Xin and Guo, Chunle and Kınlı, Furkan and Menteş, Sami and Özcan, Barış and Kıraç, Furkan and Zini, Simone and Rota, Claudio and Buzzelli, Marco and Bianco, Simone and Schettini, Raimondo and Li, Wei and Ma, Yipeng and Wang, Tao and Xu, Ruikang and Song, Fenglong and Chen, Wei-Ting and Yang, Hao-Hsiang and Huang, Zhi-Kai and Chang, Hua-En and Kuo, Sy-Yen and Liang, Zhexin and Zhou, Shangchen and Feng, Ruicheng and Li, Chongyi and Chen, Xiangyu and Song, Binbin and Zhang, Shile and Liu, Lin and Wang, Zhendong and Ryu, Dohoon and Bae, Hyokyoung and Kwon, Taesung and Desai, Chaitra and Akalwadi, Nikhil and Joshi, Amogh and Mandi, Chinmayee and Malagi, Sampada and Uppin, Akash and Sudheer Reddy, Sai and Ashok Tabib, Ramesh and Patil, Ujwala and Mudenagudi, Uma},
  booktitle={2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW)}, 
  title={NTIRE 2022 Challenge on Night Photography Rendering}, 
  year={2022},
  volume={},
  number={},
  pages={1286-1299},
  keywords={Photography;Visualization;Computer vision;Conferences;Rendering (computer graphics);Cameras;Pattern recognition},
  doi={10.1109/CVPRW56347.2022.00135},
  url={https://openaccess.thecvf.com/content/CVPR2022W/NTIRE/html/Ershov_NTIRE_2022_Challenge_on_Night_Photography_Rendering_CVPRW_2022_paper.html},
}

@article{https://doi.org/10.1002/col.22675,
author = {Ershov, Egor and Savchik, Alex and Semenkov, Ilya and Banić, Nikola and Koščević, Karlo and Subašić, Marko and Belokopytov, Alexander and Terekhin, Arseniy and Senshina, Daria and Nikonorov, Artem and Li, Zhihao and Qian, Yanlin and Buzzelli, Marco and Riva, Riccardo and Bianco, Simone and Schettini, Raimondo and Barron, Jonathan T. and Lončarić, Sven and Nikolaev, Dmitry},
title = {Illumination estimation challenge: The experience of the first 2 years},
journal = {Color Research \& Application},
volume = {46},
number = {4},
pages = {705-718},
keywords = {challenge, color constancy, illumination estimation, mixed illumination, multiple illumination, white balancing},
doi = {https://doi.org/10.1002/col.22675},
url = {https://doi.org/10.1002/col.22675},
eprint = {https://onlinelibrary.wiley.com/doi/pdf/10.1002/col.22675},
abstract = {Abstract Illumination estimation is the essential step of computational color constancy, one of the core parts of various image processing pipelines of modern digital cameras. Having an accurate and reliable illumination estimation is important for reducing the illumination influence on the image colors. To motivate the generation of new ideas and the development of new algorithms in this field, two challenges on illumination estimation were conducted. The main advantage of testing a method on a challenge over testing it on some of the known datasets is the fact that the ground-truth illuminations for the challenge test images are unknown up until the results have been submitted, which prevents any potential hyperparameter tuning that may be biased. The First illumination estimation challenge (IEC\#1) had only a single task, global illumination estimation. The second illumination estimation challenge (IEC\#2) was enriched with two additional tracks that encompassed indoor and two-illuminant illumination estimation. Other main features of it are a new large dataset of images (about 5000) taken with the same camera sensor model, a manual markup accompanying each image, diverse content with scenes taken in numerous countries under a huge variety of illuminations extracted by using the SpyderCube calibration object, and a contest-like markup for the images from the Cube++ dataset. This article focuses on the description of the past two challenges, algorithms which won in each track, and the conclusions that were drawn based on the results obtained during the first and second challenge that can be useful for similar future developments.},
year = {2021}
}

@ARTICLE{9296220,
  author={Ershov, Egor and Savchik, Alexey and Semenkov, Illya and Banić, Nikola and Belokopytov, Alexander and Senshina, Daria and Koščević, Karlo and Subašić, Marko and Lončarić, Sven},
  journal={IEEE Access}, 
  title={The Cube++ Illumination Estimation Dataset}, 
  year={2020},
  volume={8},
  number={},
  pages={227511-227527},
  keywords={Lighting;Estimation;Image color analysis;Light sources;Calibration;Semantics;Publishing;Color constancy;dataset;illumination estimation;white balancing;multiple illumination;mixed illumination},
  doi={10.1109/ACCESS.2020.3045066}}