publications

2024

1. LLM Fine-Tuning with Distributional Feedback

   E Koroleva, and S Mikhailova

   preprint, 2024

   Abs PDF

   Fine-tuning large language models (LLMs) to align with human preferences is a critical challenge in AI alignment. Current approaches predominantly rely on Reinforcement Learning from Human Feed- back (RLHF), which typically converts complex human judgments into point estimates for rewards. In this paper, we propose Distributional Feedback Learning (DFL), a novel fine-tuning approach that captures the inherent uncertainty and distribution of human preferences rather than reducing them to singular values. Our method represents human feedback as probability distributions over preference scores, allowing models to better capture preference ambiguity and variation across evaluators. We develop a distributional loss function that encourages models to predict the full distribution of human preferences rather than just their mean. Experiments across diverse tasks show that DFL outperforms standard RLHF in alignment quality, reducing catastrophic response rate by 27.3% and achieving a 18.6% improvement in preference matching on ambiguous queries. Furthermore, we demonstrate that DFL models exhibit improved calibration regarding their uncertainty, producing more nuanced outputs in scenarios where human preferences are genuinely divided. Our approach represents a significant advancement in LLM alignment techniques that better reflects the multifaceted nature of human value judgments.

2023

1. 

   DnyRNA Transformer for RNA Structure and Ractivity Profile Prediction

   E Koroleva, and S Mikhailova

   preprint, 2023

   Abs PDF

   RNA structure prediction is a fundamental challenge in molecular biology with significant impli- cations for medicine, biotechnology, and our understanding of life. In this paper, we present an enhanced transformer-based architecture for predicting RNA structure and chemical reactivity pro- files. Our approach integrates base pair probability matrices (BPPMs) with sequence information through a novel convolutional-attention mechanism and incorporates dynamic positional bias to better generalize to sequences of varying lengths. We introduce a Squeeze-and-Excitation enhancement to convolutional blocks that improves feature extraction from BPPMs and develop a specialized model for cross-reactivity prediction. Our ensemble approach achieves a mean absolute error (MAE) of 0.0626 on the RNA reactivity dataset, representing a significant improvement over existing methods. We analyze the contribution of various architectural components and demonstrate that our approach effectively captures the complex interactions between nucleotides that determine RNA structure. The proposed model has potential applications in RNA-based drug design, understanding genetic diseases, and developing novel therapeutics.

2020

1. Adaptive Feature Fusion For Multi-Camera Human Tracking

   M Sorokin, E Koroleva, A Shuvalov, and 1 more author

   preprint, 2020

   Abs PDF

   Multi-camera human tracking in crowded environments remains a challenging problem due to occlusions, illumination changes, and appearance variations across different camera views. This paper presents a novel framework that leverages adaptive feature fusion and temporal consistency constraints to improve tracking performance in complex environments. Our approach combines appearance, motion, and spatial-temporal features through a dynamic weighting mechanism that adapts to the complexity of the scene. We introduce a Confidence-Aware Association (CAA) algorithm that explicitly models tracking uncertainty and uses it to guide the data association process. Extensive experiments on three public datasets demonstrate that our method achieves comparable or superior performance to state-of-the-art approaches, with notable improvements in crowded scenes with frequent occlusions. The proposed framework achieves a MOTA score of 76.8% on the WILDTRACK dataset and 68.3% on the CAMPUS dataset, representing a 2.6% and 3.1% improvement over baseline methods, respectively. Our ablation studies highlight the effectiveness of the adaptive feature fusion mechanism and temporal consistency constraints in improving tracking robustness and accuracy.

2019

1. 

   Automated Processing and Analysis of Dermal Injury Healing Quality

   E Semenova, O Gerasimov, and E Koroleva

   In Biomechanics in Medicine and Biology, 2019

   Abs DOI PDF

   Automation in the analysis of biological data can enhance the accuracy of results and reduce processing time. Analyzing microscopic images is a common task in biology. To demonstrate the proposed method, we analyzed collagen structures in dermal tissue images. A methodology for automatic analysis of microscope images is presented. The object of analysis can be identified using a color vector. The image is then binarized and meshed. For each mesh element, the distribution of mean intercept length (MIL) is calculated. The orientation of objects is estimated using MIL approximation. An equation is proposed to assess the quality of collagen recovery. This method was applied to three sample groups: no ficin (N), ficin-treated (F), and immobilized ficin (Fi). We analyzed 10 images per group and obtained results using the described technique. Collagen recovery quality was: N group – 48% ± 8%, F group – 78% ± 7%, and Fi group – 68% ± 9%. These results suggest that ficin positively influences dermal healing. The obtained data are consistent with previously published studies.

2018

1. 

   Construction of an Inhomogeneous Finite Element Model Using CT Data

   O Sachenkov, O Gerasimov, E Koroleva, and 1 more author

   Russian Journal of Biomechanics, 2018

   Abs DOI PDF

   This study presents a methodology for constructing a finite element model (FEM) based on computed tomography (CT) data. The approach was applied to model the femur, emphasizing the importance of capturing the spatial distribution of mechanical properties in bone tissue and the need for individualized modeling. Finite element simulations were performed using ANSYS, and CT data were processed in Avizo. The analysis considered a linear inhomogeneous elastic material. Power-law relationships between optical density and mechanical properties—specifically Young’s modulus and ultimate stress—were employed. Optical density was derived from linear correlations with Hounsfield units. Mechanical properties were assigned to each finite element based on CT data. After solving the stress-strain problem, a safety factor was calculated for each node, taking into account the local material properties. Both homogeneous and inhomogeneous models were developed and analyzed. The results clearly demonstrate significant differences in stress-strain responses between the two models. The inhomogeneous model enables evaluation of local bone strength, accounting for patient-specific characteristics.

2017

1. 

   Evaluation of Mechanical Properties of a Heterogeneous Porous Structure

   O Gerasimov, E Koroleva, and O Sachenkov

   IOP Conference Series: Materials Science and Engineering, Jun 2017

   Abs DOI PDF

   This study addresses the problem of determining the macroscopic mechanical properties of porous materials based on their internal structure. The structure was characterized using the fabric tensor and porosity. An experimental investigation was carried out using a two-component cold-curing liquid polyurethane plastic (Lasilcast Lc-12). The fabricated samples were scanned using computed tomography, and the resulting data were analyzed to extract structural characteristics. A representative subvolume was selected for further analysis. Mechanical testing was then conducted to determine the fabric tensor, porosity, Young’s modulus, and Poisson’s ratio of the samples. Results are presented for several specimens. Furthermore, we examined how macroscopic mechanical properties depend on the nature of the porous structure, particularly considering porosity variation. To evaluate this dependence, we established relationships between Young’s modulus and Poisson’s ratio and two structural parameters: pore orientation angle (α) and pore ellipticity (λ). The sensitivity of deformation behavior to elastic constants was also assessed.

2. Mechanical properties and structure of bone tissue are changed after unloading handig

   T Baltina, O Sachenkov, N Ahmetov, and 3 more authors

   In Osteoporosis international, Jun 2017