All issues

In today’s technology-driven world, lite devices like Internet of Things (IoT) devices and microcontrollers (MCUs) are becoming increasingly common. These devices are more energyefficient and affordable, often with reduced features compared to the standard versions such as very limited memory and processing power for typical machine learning models. However, modern machine learning models can have millions of parameters, resulting in a large memory footprint. This complexity not only makes it difficult to deploy these large models on resource constrained devices but also increases the risk of latency and inefficiency in processing, which is crucial in some cases where real-time responses are required such as autonomous driving and medical diagnostics. In recent years, neural networks have seen significant advancements in model optimization techniques that help deployment and inference on these small devices. This narrative review offers a thorough examination of the progression and latest developments in neural network optimization, focusing on key areas such as quantization, pruning, knowledge distillation, and neural architecture search. It examines how these algorithmic solutions have progressed and how new approaches have improved upon the existing techniques making neural networks more efficient. This review is designed for machine learning researchers, practitioners, and engineers who may be unfamiliar with these methods but wish to explore the available techniques. It highlights ongoing research in optimizing networks for achieving better performance, lowering energy consumption, and enabling faster training times, all of which play an important role in the continued scalability of neural networks. Additionally, it identifies gaps in current research and provides a foundation for future studies, aiming to enhance the applicability and effectiveness of existing optimization strategies.

The key problems of remote forest pathology monitoring for coniferous forests affected by insect pests have been analyzed. It has been demonstrated that addressing these tasks requires the use of multiclass classification results for coniferous trees in high- and ultra-high-resolution images, which are promptly obtained through monitoring via satellites or unmanned aerial vehicles (UAVs). An analytical review of modern models and methods for multiclass classification of coniferous forest images was conducted, leading to the development of three fully convolutional neural network models: Mo-U-Net, At-Mo-U-Net, and Res-Mo-U-Net, all based on the classical U-Net architecture. Additionally, the Segformer transformer model was modified to suit the task. For RGB images of fir trees Abies sibirica affected by the four-eyed bark beetle Polygraphus proximus, captured using a UAV-mounted camera, two datasets were created: the first dataset contains image fragments and their corresponding reference segmentation masks sized 256 × 256 × 3 pixels, while the second dataset contains fragments sized 480 × 480 × 3 pixels. Comprehensive studies were conducted on each trained neural network model to evaluate both classification accuracy for assessing the degree of damage (health status) of Abies sibirica trees and computation speed using test datasets from each set. The results revealed that for fragments sized 256 × 256 × 3 pixels, the At-Mo-U-Net model with an attention mechanism is preferred alongside the Modified Segformer model. For fragments sized 480 × 480 × 3 pixels, the Res-Mo-U-Net hybrid model with residual blocks demonstrated superior performance. Based on classification accuracy and computation speed results for each developed model, it was concluded that, for production-scale multiclass classification of affected fir trees, the Res-Mo-U-Net model is the most suitable choice. This model strikes a balance between high classification accuracy and fast computation speed, meeting conflicting requirements effectively.

The paper presents a hybrid artificial neural network model based on an architecture that incorporates a convolutional image encoder (CNN) and an attention module (Attention-based Multiple Instance Learning, Attention MIL). This module aggregates informative features from a sequence of frames capturing the flame spraying process. Additional technological parameters—air pressure, propane pressure, and standoff distance — are integrated into the model via a tabular channel, enabling it to account for the relationship between visual data and numerical process regime characteristics. The software implementation was developed using the Streamlit platform and the PyTorch library. It features an interactive interface for model training and result visualization, analysis of attention weights across frames, and a prediction mode for output characteristics: surface roughness ($R_a$) and the mass of the deposited coating ($m$). Experimental studies were conducted on data from real-world technological processes, and a comparative analysis of the accuracy of various model configurations was performed. The results demonstrate that the hybrid neural network, which combines visual and tabular features, achieves higher prediction accuracy compared to models using only a single modality. Furthermore, when comparing different implementations of the hybrid network, it was established that using the attention mechanism to process the series of flame spray images provides a significant increase in accuracy over a simple averaging of features without attention. The application includes an attention visualization module that creates a montage of the most significant frames and displays their attention weights, allowing users to identify which frames had the greatest influence on the prediction. The model’s capability for export to the ONNX format for integration into process control systems is also demonstrated. The proposed approach showcases the effectiveness of fusing visual and tabular information for manufacturing process monitoring tasks. The model can serve as a foundation for developing a decision support system or an automated quality control system for coatings produced by flame spraying. The limitations of the implemented model and prospects for its further development are also considered.

In this paper, we propose an adaptive algorithm that simulates the process of forming the initial behavioral skills on the example of the system ‘eye-arm’ animat. The situation is the formation of the initial behavioral skills occurs, for example, when a child masters the management of their hands by understanding the relationship between baseline unidentified spots on the retina of his eye and the position of the real object. Since the body control skills are not ‘hardcoded’ initially in the brain and the spinal cord at the level of instincts, the human child, like most young of other mammals, it is necessary to develop these skills in search behavior mode. Exploratory behavior begins with trial and error and then its contribution is gradually reduced as the development of the body and its environment. Since the correct behavior patterns at this stage of development of the organism does not exist for now, then the only way to select the right skills is a positive reinforcement to achieve the objective. A key feature of the proposed algorithm is to fix in the imprinting mode, only the final action that led to success, and that is very important, led to the familiar imprinted situation clearly leads to success. Over time, the continuous chain is lengthened right action — maximum use of previous positive experiences and negative ‘forgotten’ and not used.

Thus there is the gradual replacement of the random search purposeful actions that observed in the real young. Thus, the algorithm is able to establish a correspondence between the laws of the world and the ‘inner feelings’, the internal state of the animat. The proposed animat model was used 2 types of neural networks: 1) neural network NET1 to the input current which is fed to the position of the brush arms and the target point, and the output of motor commands, directing ‘brush’ manipulator animat to the target point; 2) neural network NET2 is received at the input of target coordinates and the current coordinates of the ‘brush’ and the output value is formed likelihood that the animat already ‘know’ this situation, and he ‘knows’ how to react to it. With this architecture at the animat has to rely on the ‘experience’ of neural networks to recognize situations where the response from NET2 network of close to 1, and on the other hand, run a random search, when the experience of functioning in this area of the visual field in animat not (response NET2 close to 0).

Automated urban traffic monitoring systems are widely used to solve various tasks in intelligent transport systems of different regions. They include video enforcement, video surveillance, traffic management system, etc. Effective traffic management and rapid response to traffic incidents require continuous monitoring and analysis of information from these complexes, as well as time series forecasting for further anomaly detection in traffic flow. To increase the forecasting quality, data fusion from different sources is needed. It will reduce the forecasting error, related to possible incorrect values and data gaps. We implemented the approach for short-term and middle-term forecasting of traffic flow (5, 10, 15 min) based on data fusion from video enforcement and video surveillance systems. We made forecasting using different recurrent neural network architectures: LSTM, GRU, and bidirectional LSTM with one and two layers. We investigated the forecasting quality of bidirectional LSTM with 64 and 128 neurons in hidden layers. The input window size (1, 4, 12, 24, 48) was investigated. The RMSE value was used as a forecasting error. We got minimum RMSE = 0.032405 for basic LSTM with 64 neurons in the hidden layer and window size = 24.

Deep learning’s power stems from complex architectures; however, these can lead to overfitting, where models memorize training data and fail to generalize to unseen examples. This paper proposes a novel probabilistic approach to mitigate this issue. We introduce two key elements: Truncated Log-Uniform Prior and Truncated Log-Normal Variational Approximation, and Automatic Relevance Determination (ARD) with Bayesian Deep Neural Networks (BDNNs). Within the probabilistic framework, we employ a specially designed truncated log-uniform prior for noise. This prior acts as a regularizer, guiding the learning process towards simpler solutions and reducing overfitting. Additionally, a truncated log-normal variational approximation is used for efficient handling of the complex probability distributions inherent in deep learning models. ARD automatically identifies and removes irrelevant features or weights within a model. By integrating ARD with BDNNs, where weights have a probability distribution, we achieve a variational bound similar to the popular variational dropout technique. Dropout randomly drops neurons during training, encouraging the model not to rely heavily on any single feature. Our approach with ARD achieves similar benefits without the randomness of dropout, potentially leading to more stable training.

To evaluate our approach, we have tested the model on two datasets: the Canadian Institute For Advanced Research (CIFAR-10) for image classification and a dataset of Macroscopic Images of Wood, which is compiled from multiple macroscopic images of wood datasets. Our method is applied to established architectures like Visual Geometry Group (VGG) and Residual Network (ResNet). The results demonstrate significant improvements. The model reduced overfitting while maintaining, or even improving, the accuracy of the network’s predictions on classification tasks. This validates the effectiveness of our approach in enhancing the performance and generalization capabilities of deep learning models.

Considered the application of physics-informed neural networks using multi layer perceptrons to solve Cauchy initial value problems in which the right-hand sides of the equation are continuous monotonically increasing, decreasing or oscillating functions. With the use of the computational experiments the influence of the construction of the approximate neural network solution, neural network structure, optimization algorithm and software implementation means on the learning process and the accuracy of the obtained solution is studied. The analysis of the efficiency of the most frequently used machine learning frameworks in software development with the programming languages Python and C# is carried out. It is shown that the use of C# language allows to reduce the time of neural networks training by 20–40%. The choice of different activation functions affects the learning process and the accuracy of the approximate solution. The most effective functions in the considered problems are sigmoid and hyperbolic tangent. The minimum of the loss function is achieved at the certain number of neurons of the hidden layer of a single-layer neural network for a fixed training time of the neural network model. It’s also mentioned that the complication of the network structure increasing the number of neurons does not improve the training results. At the same time, the size of the grid step between the points of the training sample, providing a minimum of the loss function, is almost the same for the considered Cauchy problems. Training single-layer neural networks, the Adam method and its modifications are the most effective to solve the optimization problems. Additionally, the application of twoand three-layer neural networks is considered. It is shown that in these cases it is reasonable to use the LBFGS algorithm, which, in comparison with the Adam method, in some cases requires much shorter training time achieving the same solution accuracy. The specificity of neural network training for Cauchy problems in which the solution is an oscillating function with monotonically decreasing amplitude is also investigated. For these problems, it is necessary to construct a neural network solution with variable weight coefficient rather than with constant one, which improves the solution in the grid cells located near by the end point of the solution interval.

For cognitive radio systems, it is important to use efficient algorithms that search for free channels that can be provided to secondary users. Therefore, this paper is devoted to improving the accuracy of prediction frequency resource occupancy of a cellular communication system using spatiotemporal radio environment maps. The formation of a radio environment map is implemented for the fourthgeneration cellular communication system Long-Term Evolution. Taking this into account, a model structure has been developed that includes data generation and allows training and testing of an artificial neural network to predict the occupancy of frequency resources presented as the contents of radio environment map cells. A method for assessing prediction accuracy is described. The simulation model of the cellular communication system is implemented in the MatLab. The developed frequency resource occupancy prediction model is implemented in the Python. The complete file structure of the model is presented. The experiments were performed using artificial neural networks based on the Long Short-Term Memory and Kolmogorov – Arnold neural network architectures, taking into account its modification. It was found that with an equal number of parameters, the Kolmogorov –Arnold neural network learns faster for a given task. The obtained research results indicate an increase in the accuracy of prediction the occupancy of the frequency resource of the cellular communication system when using the Kolmogorov – Arnold neural network.

This paper considers the inverse problem of seismic exploration — determining the structure of the media based on the recorded wave response from it. Large cracks are considered as target objects, whose size and position are to be determined.

he direct problem is solved using the grid-characteristic method. The method allows using physically based algorithms for calculating outer boundaries of the region and contact boundaries inside the region. The crack is assumed to be thin, a special condition on the crack borders is used to describe the crack.

The inverse problem is solved using convolutional neural networks. The input data of the neural network are seismograms interpreted as images. The output data are masks describing the medium on a structured grid. Each element of such a grid belongs to one of two classes — either an element of a continuous geological massif, or an element through which a crack passes. This approach allows us to consider a medium with an unknown number of cracks.

The neural network is trained using only samples with one crack. The final testing of the trained network is performed using additional samples with several cracks. These samples are not involved in the training process. The purpose of testing under such conditions is to verify that the trained network has sufficient generality, recognizes signs of a crack in the signal, and does not suffer from overtraining on samples with a single crack in the media.

The paper shows that a convolutional network trained on samples with a single crack can be used to process data with multiple cracks. The networks detects fairly small cracks at great depths if they are sufficiently spatially separated from each other. In this case their wave responses are clearly distinguishable on the seismogram and can be interpreted by the neural network. If the cracks are close to each other, artifacts and interpretation errors may occur. This is due to the fact that on the seismogram the wave responses of close cracks merge. This cause the network to interpret several cracks located nearby as one. It should be noted that a similar error would most likely be made by a human during manual interpretation of the data. The paper provides examples of some such artifacts, distortions and recognition errors.

The paper presents a cascade model of a physically supported neural network designed to predict pressure drop in three-phase flow (oil, gas, water) in a pipe section with various angles of inclination. To overcome the constraints of existing empirical correlations and computation-intensive numerical modeling methods, we propose an architecture that decomposes the problem into three sequential physically interpretable subtasks: regression prediction of the fluid hold-up coefficient, fluid flow regime classification, and pressure gradient evaluation. Each subtask is solved by a separate fully connected neural network, the output of which is passed to the next model in the cascade. Training and testing of the proposed architecture was performed on an extensive synthetic dataset (8 · 10⁷ records) generated using a semi-empirical model. Verification is performed on independent experimental data. A comparative analysis with a single fully connected (non-cascade) neural network is made, and the sensitivity of the models is examined using Sobol and Borgonovo methods. The cascade model demonstrates superior accuracy and ensures high interpretability of results by providing intermediate physical parameters (fluid hold-up coefficient, flow regime). The developed model has low computational complexity, which allows it to be used in real-time systems and digital twins of hydraulic systems in the oil and gas industry.