This article presents a time-domain speech separation model TIGER-DFT based on a TIGER core for separation part with encoder-decoder layers that mimics a windowed Discrete Fourier transformation. Our proposed model has achieved a separation performance of 7.72 dB in SI-SDR and 9.65 dB in SI-SDRi for noise Libri2Mix dataset outperformed original TIGER model for 0.4 dB for both metrics.
This paper considers the problem of adaptive optimization of the training dataset for neural networks in image classification tasks. It has been shown that using the full training dataset is not the best solution. Different training samples have different value for the model. Some samples are representative, some are important for class boundaries, some preserve diversity, and some may be noisy or suspicious. A method for adaptive formation of the active training subset has been proposed. Unlike traditional methods, the proposed approach evaluates the role of each sample during training.
The article considers methods and means of face recognition in a video data stream using elements of machine learning. The main approaches to face identification based on computer vision methods are analyzed, in particular the principal component analysis (PCA) method, the local binary pattern analysis (LBP) method and the linear discriminant analysis (LDA, Fisherfaces) method. The principles of facial feature formation and algorithms for their classification are described.
This paper investigates modern approaches to the application of artificial intelligence (AI) models in prosthetic systems from the perspective of information technologies. The aim of the study is to provide a systematic analysis and comparative evaluation of contemporary AI models used in prosthetics, with a focus on control algorithms, biosignal processing methods, and the integration of intelligent solutions into real-time prosthetic devices.
The paper presents an integrated framework for dynamic routing of unmanned aerial vehicles (UAVs) operating in, uncertain, and rapidly changing environments. After reviewing classical deterministic, sampling complex -based, bio-inspired, machine learning, and hybrid path-planning methods, their limitations with respect to real-time replanning, energy efficiency, and scalability to multi-UAV missions are critically analysed.
Summary. Automation of database indexing is a crucial component of modern database management systems that enhances search performance, scalability, and relevance in large-scale data environments. This paper explores the application of deep learning algorithms for building and optimizing vector indexes capable of automatic adaptation to changes in data structure and query patterns. An experimental comparison was conducted between traditional indexing methods (B-Tree, GIN in PostgreSQL) and vector-based indexing using Sentence-BERT embeddings implemented in FAISS and Milvus systems.
This paper examines the problem of predicting a user’s multidimensional emotional state from textual records under conditions where most existing text-based approaches emphasize either categorical emotion recognition or coarse sentiment polarity, which limits the interpretability of broader affective assessment.
The extremely pressing issue of breast cancer diagnosis, which remains one of the leading causes of mortality, requires innovative approaches to improve the accuracy and speed of cancer diagnosis.
This article presents a hybrid speech separation model designed for efficient deployment on edge devices, focusing on optimizing both performance and computational resources. This study proposes a novel hybrid architecture that combines the strengths of Conv-TasNet and SuDoRM- RF models, leveraging their fully-convolutional structures to achieve efficient separation with minimal resource usage. The proposed model has obtained a separation performance of 10.59 db in SI-SDRi for clean Libri2Mix dataset for only 1.17 M parameters with only 0.92 GMACs/s.
The paper considers methods for optimizing training samples for deep learning algorithms through the use of random point processes, such as Matern of the first and second types, Gibbs, Gaussian, and Poisson processes. An approach to reducing training data without sacrificing informativeness is proposed, enabling a decrease in computational costs and mitigating the risk of overfitting.