15-17 червня студенти другого року навчання магістерської програми «Науки про дані» будуть захищати свої дипломні роботи. Запрошуємо долучитися до перегляду захистів робіт онлайн. Розклад, теми та опис робіт, а також посилання на YouTube трансляції для окремих днів наводяться нижче.
- 10:00, Iaroslav Plutenko, Incorporating Metadata for Semantic Segmentation by Employing Channel Attention Mechanism
- The meta-information accompanying data from image acquisition devices has limited use in microscopy image processing techniques involving Deep Learning. This project aims to incorporate the supplementary metadata for semantic segmentation by employing a channel selection mechanism in convolutional networks outlining its potential benefits and practical applications where metadata can be used for switching tasks within a master model. The results of conducted experiments show that meta-information is helpful, and the phenomenon is more expressed with incompatible segmentation tasks, where a multi-head model or separate models are required otherwise. Overall, we have achieved a slight increase in scores for similar tasks as well and demonstrated the applicability of the CNN model for separate tasks, forcing it to work as an ensemble, leveraging the beneficial effect of multi-task learning.
- 10:40, Andrii Blagodyr, Raspberry quality detection in visual spectrum using neural networks
- The thesis presents the raspberry quality detection approach based on a convolutional neural network with U-net architecture and compared with PSPnet architecture. The limited possibility to use manual labour when growing, sorting, processing vegetables, fruits and berries in the face of increasing risks of new pandemics determines the study’s relevance. For the research, a neural network of the U-net architecture has been chosen based on the narrow focus of the task and small repetitive patterns. The neural network of the U-net architecture has proven itself well in solving problems of image segmentation in biomedical researches. Therefore, the author decided to expand the scope of this tool to a new area of investigation. The research is carried out on the data that the researchers have collected for the experiment. The dataset for the experiment has been generated manually based on images of different varieties of raspberries and various states of raspberry fruits. This research is expected to become a part of the complex robotic system for solving the problem of manual berry fruits sorting.
- 11:20, Rostyslav Zatserkovnyi, 3D Head Model Estimation from a Single Photo
- Today, 3D human head models are widely used in fields such as computer vision, entertainment, healthcare, and biometrics. Since a high-quality scan of a human head is expensive and time-consuming to obtain, machine learning algorithms are used to estimate the shape and texture of a 3D model from a single “in-the-wild” photograph, often taken at extreme angles or with non-uniform illumination. However, as a full head texture cannot be trivially inferred from a single photograph due to self-occlusion, many only focus on modeling an incomplete and partially textured model of the human head. This work proposes a machine learning pipeline that reconstructs a fully textured 3D head model from a single photograph. We collect a novel dataset of 99.3 thousand high-resolution human head textures created from synthetic celebrity photographs. To the best of our knowledge, this is the first UV texture dataset of a similar scale and fidelity. Using this dataset, we train a free-form inpainting GAN that learns to recreate full head textures from partially obscured projections of the input photograph.
- 13:00, Alexander Onbysh, Point cloud human pose estimation using capsule networks
- Human pose estimation based on points cloud is an emerging field that develops with 3D scanning devices’ popularity. Build-in LiDAR technology in mobile phones and a growing VR market creates a demand for lightweight and accurate models for 3D point cloud. Widely advanced deep learning tools are mainly used for structured data, and they face new challenges in unstructured 3D space. Recent research on capsule networks proves that this type of model outperforms classical CNN architectures in tasks that require viewpoint invariance from the model. Thus capsule networks challenge multiple issues of classic CNNs like preserving the orientation and spatial relationship of extracted features, which could significantly improve the 3D points cloud classification task’s performance. The project’s objective is to experimentally assess the applicability of capsule neural network architecture to the task of point cloud human pose estimation and measure performance on non-synthetic data. Additionally, measure noise sustainability of capsule networks for 3D data compared to regular models. Compare models’ performance with restricted amount of training data.
- 13:40, Ruslan Partsey, Robust Visual Odometry for Realistic Point-Goal Navigation
- The ability to navigate in complex environments is a fundamental skill of a home robot. Despite extensive study, indoor navigation in unseen environments under noisy actuation and sensing and without access to precise localization continues to be an open frontier for research in Embodied AI. In this work, we focus on designing a visual odometry module for robust egomotion estimation and its integration with navigation policy for efficient navigation under noisy actuation and sensing. Specifically, we study how the observations transformations and incorporating meta-information available to the navigation agent impacts visual odometry model generalization performance. We present a set of regularization techniques that can be implemented as train- and test-time augmentations to increase the robustness to noise. Navigation agent, equipped with our visual odometry module, reaches the goal in 86% of episodes and scores 0.66 SPL in Habitat Challenge 2021 benchmark.
- 10:30, Mykola Trokhymovych, Automated Fact-checking for Wikipedia
- The incoming flow of information is continuously increasing along with the disinformation piece that can harm society. Filtering unreliable content helps keep Wikipedia as free as possible of disinformation, making it one of the most significant reliable information sources. Consequently, Wikipedia’s knowledge base is widely used for facts verification academic research. The main goal of our work is to transform recent academic achievements into a practical open-source Wikipedia-based fact-checking application that is both accurate and efficient. We review the primary NLI related datasets and study their relevant limitations. As a result, we propose the data filtering method that improves the model’s performance and generalization. We show that transfer learning for NLI models are not working well, and complete model training is needed to achieve the best result on a specific dataset. We come up with an unsupervised fine-tuning of the Masked Language model on field-specific texts for model domain adaptation. Finally, we present the new fact-checking system WikiCheck API that automatically performs a facts validation process based on the Wikipedia knowledge base. It is comparable to SOTA solutions in terms of accuracy and can be used on low memory CPU instances.
- 11:20, Maksym Tarnavskyi, Improving Sequence Tagging for Grammatical Error Correction
- In this work, we investigated the recent sequence tagging approach for the Grammatical Error Correction task. We compared the impact of different transformer-based encoders of base and large configurations and showed the influence of tags’ vocabulary size. Also, we discovered ensembling methods on data and model levels. We proposed two methods for selecting better quality data and filtering noisy data. We generated new training GEC data based on knowledge distillation from an ensemble of models and discovered strategies for its usage. Our best ensemble without pre-training on the synthetic data achieves a new SOTA result on BEA-2019 (test), in contrast, when the newest obtained results were achieved with pre-training on synthetic data.
- 13:00, Nataliia Novosad, Investment modeling of agricultural land valuation in Ukraine
- Evaluating the fair market value of land is a complicated and expensive process carried out by experts. The final valuation is the main factor for investment decision-making. The asset valuation has to include many components, particularly the estimation of the future net income, risks, and opportunities. The objective of this study is to find the fair value of the agricultural land and compare the different approaches. We consider planting four crops: wheat, maize, soybeans, and sunflower. In this project, we estimated future crop prices, yields, and expenses to predict income. We started with the simple income method and showed how the value changes when complicating the method and considering new real options like crop rotation, optimizing crop portfolio, and installing an irrigation system. Moreover, we analyze the sensitivity of the estimated value to the economic situation (discount rate and price growth rate). The data we are considering is the land in Ukraine, specifically in the Kherson region. Due to the land moratorium, the free market does not exist, and the farmland in Ukraine is usually underestimated. Thus, this causes the impossibility of direct comparison of the proposed valuation and the absence of needed open data. Therefore, we rely on expert opinion in different aspects of the project.
- 13:40, Petro Bodnar, Detecting Patterns of Coordinated News Article Dissemination
- This study aims at devising methods for detecting coordination among content spreaders at scale. We focus on methods which uncover the latent structure of the content dissemination networks from the time-series of publications. We identify the advantages of generative models, especially self-exiting stochastic processes, for modeling information cascades and detecting structural patterns in groups of events. We validate the most popular of these models – the Multivariate Hawkes processes – on a large dataset of news websites and achieve an improvement in comparison to simpler baselines, e.g., cosine similarity between time-series of publications.
- 10:00, Oleksandr Bratus, BLE Mesh Reliability Optimization using Neural Networks
- The Bluetooth Low Energy (BLE) Mesh network technology is one of the newest technologies in the wireless communication domain. Due to low cost and low power consumption, it has already become widespread and has the potential for a wide range of applications. However, the flooding algorithm on which based BLE Mesh data transmission process impacts strongly on networks reliability. Because improper network setup can be critical to ensuring sufficient network reliability, it is necessary to be able to predict the network reliability in order to be able to reconfigure the network to improve its reliability. In this master thesis, we propose neural network approaches that predict the reliability of both the entire network and its individual nodes. Presented results demonstrate that trained neural networks are scalable by providing high accuracy of predictions on networks of different sizes.
- 10:40, Sevil Smailova, Modeling and Prediction of Alzheimer’s Disease Progression
- Alzheimer’s Disease is an irreversible disease that causes a decline in cognitive abilities and leads to dementia. Many efforts are applied to understand the behavior of the disease progression and foresee its future state. The metrics that assess the level of cognition are named as cognitive scores. The dynamics of cognitive scores help understand the future disease progression. However, there is a lack of understanding on what is the best benchmark for the predicted value of the cognitive score. Moreover, there could be cases when the future value of the cognitive score is not statistically different comparing to the current value. In this work we discover those patients that by design cannot have the dynamics in their progression of cognitive scores. We justify that the dynamics of progression for Cognitively Normal patients do not change over five years. We reveal that there is no statistically significant change in progression after the 1-year follow-ups. We unified the evaluation framework of different imputation, feature selection methods and machine learning models on different time to prediction settings as well as on different patient populations.
- 11:20, Kateryna Zabava, Real-time inverse kinematics and inverse dynamics from motion capture
- This work applies machine learning to solving inverse dynamics and inverse kinematics tasks from the motion capture data. This approach may simplify the calculation process and help do scientific simulations as part of a physics engine that describes the neural control of human motion and decodes movement intent in individuals with neural damage. The existing algorithm has to be modified for every experiment and takes a significant amount of time to execute. It is also sensitive to noise and missing data, and it is not a real-time calculation. We propose a solution of inverse kinematics tasks with neural networks. Here we report accuracy results both on clean data and noisy data. We also apply a similar approach for the inverse dynamics task. The approach shows high accuracy on clean data, but this accuracy decreases if applied to the noisy data.
- 13:00, Mykhailo Manukian, Real-time simulation of arm and hand dynamics using ANN
- The physics of body dynamics is a complex problem solved by the nervous system in real-time during the planning and execution of movements. The human arm and hand have complex mechanics involving hundreds of muscles that actuate over 30 degrees of freedom (DOF). To date, the problems of this complexity remain unsolved in engineering; yet, the nervous system computes control signals in a robust, accurate, and time-efficient manner. Neuroprosthetics require similar computations for the decoding of intent and encoding of sensory feedback. The trade-off of required computational accuracy and latency is hard to resolve with classical physics; thus, this research aims to develop “good-enough” approximations of these computations using machine learning methods, such as artificial neural networks (ANN). The kinematic and kinetic temporal computations that rely on the diverse number of terms within the equations of motion are consistent with the recurrent neural network (RNN) architectures. This study will test the general hypothesis that the inverse dynamics of arm and hand can be captured with RNN formulation and explore the utility of different architectures: i) simple Recurrent ANN, ii) Gated Recurrent Unit (GRU) ANN, and iii) Long Short-Term Memory (LSTM) ANN. The inverse problem is the mapping from joint kinematics (position, velocity, acceleration) to joint kinetics (torque). The training and testing datasets were derived from the physical model of arm and hand performing point-to-point movements between realistic postures arranged in a grid within the physiological range of motion. Lastly, we assessed the execution latency of the machine learning solutions in the context of real-time requirements for prosthetic applications.
- 13:40, Yuriy Pryyma, Central pattern generator model using Spiking neural networks
- Locomotor control involves dynamic mechanical and neural interactions that are essential for survival. The neural locomotor pathways contain the central pattern generator (CPG), a network of neurons embedded into the spinal cord and generating dynamic output for walking and running. Even though there are multiple formulations of the CPG, from coupled oscillators to complex networks of Hodgkin-Huxley neurons, the optimal choice of model implementation depends on its application. The choice of a formulation is often described as the trade-off between complexity and the level of details in the model’s function. However, the advantages between different formulations have not been established. Recently, the spiking neural networks (SNN) have gained popularity as a biological analog for neural dynamics that uses methodology developed for artificial neural networks. This formulation uses spiking frequency instead of rate signals to accomplish dynamic computations with the integrate-and-fire neurons. In this study, we aimed to create the framework for comparing a versatile CPG rate model and its implementation with the model build with SNN. We used a neuromorphic software package (Nengo) to develop and validate a bilateral CPG model’s structural and functional details based on the half-center oscillators. The spiking model shows similar precision for calculating the empirical phase-duration characteristics of gait in cats as the rate model, and it also reproduces the linear relationship between the CPG input and the empirical limb speed of forward progression. While the phase characteristic was used to optimize neural dynamics, the input relationship with the limb speed is the product of the model structure. Furthermore, the spiking model has increased tolerance to temporal noise, and it can withstand some structural damage. The spiking and rate models require further comparative analysis. Overall, the development of adaptable spiking models could help integrate the biomimetic components within the control systems for assistive robotics and electrical stimulation devices to rehabilitate locomotion after central and peripheral injuries.