Methods of machine learning and data mining are becoming the cornerstone in information technologies with real-time image and video recognition methods getting more and more attention. While computational system architectures are getting larger and more complex, their learning methods call for changes, as training datasets often reach tens and hundreds of thousands of samples, therefore increasing the learning time of such systems. It is possible to reduce computational costs by tuning the system structure to allow fast, high accuracy learning algorithms to be applied. This paper proposes a system based on extended multidimensional neo-fuzzy units and its learning algorithm designed for data streams processing tasks. The proposed learning algorithm, based on the information entropy criterion, has significantly improved the system approximating capabilities. Experiments have confirmed the efficiency of the proposed system in solving real-time video stream recognition tasks.
15 December 2020 in Information Technology and Management Science
The probabilistic neuro-fuzzy system to solve the image classification-recognition task is proposed. The considered system is a “hybrid” of Specht’s probabilistic neural network and the neuro-fuzzy system of Takagi-Sugeno-Kang. It is designed to solve tasks in case of overlapping classes. Also, it is supposed that the initial data that are fed on the input of the system can be represented in numerical, rank, and nominal (binary) scales. The tuning of the network is implemented with the modified procedure of lazy learning based on the concept “neurons at data points”. Such a learning approach allows substantially reducing the consumption of time and does not require large amounts of training dataset. The proposed system is easy in computational implementation and characterised by a high classification speed, as well as allows processing information both in batch and online mode.
1 January 2022 in Mathematical Biosciences and Engineering
In the paper, we propose the modified generalized neo-fuzzy system. It is designed to solve the pattern-image recognition task by working with data that are fed to the system in the image form. The neo-fuzzy system can work with small training datasets, where classes can overlap in a features space. The core of the system under consideration is a modification of multidimensional generalized neuro-fuzzy neuron with an additional softmax activation function in the output layer instead of the defuzzification layer and quartic-kernel functions as membership ones. The learning procedure of the system combined cross-entropy criterion optimization using a matrix version of the optimal by speed Kaczmarz-Widrow-Hoff algorithm with the additional filtering (smoothing) properties. In comparison to the well-known systems, the modified neo-fuzzy one provides both numerical and computational implementation simplicity. The computational experiments have proved the effectiveness of the modified generalized neo-fuzzy-neuron, including the situation with shot training datasets.
19 November 2021 in The Open Bioinformatics Journal
Background: The medical diagnostic task in conditions of the limited dataset and overlapping classes is considered. Such limitations happen quite often in real-world tasks. The lack of long training datasets during solving real tasks in the problem of medical diagnostics causes not being able to use the mathematical apparatus of deep learning. Additionally, considering other factors, such as in a dataset, classes can be overlapped in the feature space; also data can be specified in various scales: in the numerical interval, numerical ratios, ordinal (rank), nominal and binary, which does not allow the use of known neural networks. In order to overcome arising restrictions and problems, a hybrid neuro-fuzzy system based on a probabilistic neural network and adaptive neuro-fuzzy interference system that allows solving the task in these situations is proposed. Methods: Computational intelligence, artificial neural networks, neuro-fuzzy systems compared to conventional artificial neural networks, the proposed system requires significantly less training time, and in comparison with neuro-fuzzy systems, it contains significantly fewer membership functions in the fuzzification layer. The hybrid learning algorithm for the system under consideration based on self-learning according to the principle “Winner takes all” and lazy learning according to the principle “Neurons at data points” has been introduced. Results: The proposed system solves the problem of classification in conditions of overlapping classes with the calculation of the membership levels of the formed diagnosis to various possible classes. Conclusion: The proposed system is quite simple in its numerical implementation, characterized by a high speed of information processing, both in the learning process and in the decision-making process; it easily adapts to situations when the number of diagnostics features changes during the system's functioning.
26 September 2023 in EAI Endorsed Transactions on Pervasive Health and Technology
INTRODUCTION: The task of approximation of complex nonlinear dependencies, especially in the case of short datasets, is important in various applied fields of medicine. Global approximation methods describe the generalized behavior of the model, while local methods explain the behavior of the model at specific data points. Global-local approximation combines both approaches, which makes such methods a powerful tool for processing short sets of medical data that can have both broad trends and local variations. OBJECTIVES: This paper aims to improve the method of sequential obtaining global and local components of the response surface to increase the accuracy of prediction in the case of short sets of medical data. METHODS: In this paper, the authors developed a new method that combined two ANNs: a non-iterative SGTM neural-like structure for obtaining the global component and GRNN as a powerful tool of local approximation in the case of short datasets. RESULTS: The authors have improved the method of global-local approximation due to the use of a General Regression Neural Network instead of RBF ANN for obtaining the local component, which ensured an increase in the accuracy of the body fat prediction task. The authors optimized the operation of the method and investigated the efficiency of the sequential obtaining global and local components of the response surface in comparison with the efficiency using a number of existing methods. CONCLUSION: The conducted experimental studies for solving the body fat prediction task showed the high efficiency of using the improved method in comparison with a number of existing methods, including ensemble methods.