Performance evaluation and optimization of a deep learning parameterization method trained from submesoscale-permitting ocean simulations

In this paper, we propose a neural network-based scalable fast intra prediction algorithm in H.264 in order to reduce redundant calculation time by selecting the best mode of 4 × 4 and 16 × 16 intra prediction. In this reason, it is possible to encode compulsively by 4 × 4 intra prediction mode for current MB(macro block)'s best prediction mode without redundant mode decision calculation in accordance with neural network's output resulted from corelation of adjacent encoded four left, up-left, up and up-right blocks. If there is any one of MBs encoded by 16 × 16 intra prediction among four MBs adjacent to current MB, the probability of re-prediction into 16 × 16 intra prediction will become high. We can apply neural networks in order to decide whether to force into 4 × 4 intra prediction mode or not. We can also control both the bit rates and calculation time by modulating refresh factors and weights of neural network's output depend on error back-propagation, which is called refreshing. In case of encoding several video sequences by the proposed algorithm, the total encoding time of 30 input I frames are reduced by 20% ∼ 65% depending upon the test vector compared with JM 8.4 by using neural networks and by modulating scalable refreshing factor. On the other hand, total encoding bits are increased by 0.8% ∼ 2.0% at the cost of reduced SNR of 0.01 dB.

51 - Techniques and Applications of Neural Networks for Fuzzy Rule Approximation

The main contribution of this paper is to develop a method, using the Newton-Raphson me/hod, to search for the unknown part of the inputs of a multilayer neural network with given outputs and known inputs. To use the Newton-Raphson method, a method of expressing a jacobian by neural network parameters is developed first. A locally linearized relation between inputs and outputs of neural network is then derived. With this, iterative Newton-Raphson searches are performed until satisfactory results are obtained. The method shows rapid convergence, compared with previous approaches. While deriving the inverse of the neural network, some types of optimality, which are problem dependent, are resolved. Although the method shows fast convergence, this type of solution yields some inversion error due to the neural network modelling error. The second contribution of this paper is to propose a novel structure which can eliminate the inversion error caused by the neural network modelling error. The proposed method has a simple structure, but shows good performance as it has a feedforward structure and other beneficial features. Through computer experiments, the proposed methods show good performances in solving inverse kinematics of redundant robots and controlling nonlinear plant.

Relevance. When solving problems of stabilization of dynamic objects, classical model predictive control is widely used. It provides high quality control by solving the optimization problem at each step, but it has significant computing costs, which limits its application in real-time systems with high requirements for update frequency. Therefore, the question of investigating the applicability of a neural network regulator trained on a model predictive regulator (MPC) when solving the problem of stabilizing the position of a dynamic object with a limited computational and time resource is relevant. Goal.  The purpose of the presented work was to develop and study a neural network regulator trained on the basis of an MPC regulator to stabilize the position of a dynamic object on a mobile platform. Methods. When performing the work, methods of system analysis, simulation modeling, as well as experimental tests on the bench were used. Results and conclusions. As part of the study, a neural network regulator was developed and trained that approximates the behavior of MPC based on data obtained when controlling a real balancing platform. The training was conducted on the input and output data of the MPC without using the internal model of the system, which made it possible to reproduce the dynamics of the regulator at significantly lower computational costs. Experimental results showed that the neural network model provides a stabilization quality comparable to the original MPC, while the calculation time was reduced from 47 ms to 1.6 ms, which amounted to an acceleration value of 29 times. The proposed approach demonstrates the potential of neural network control methods in the problems of replacing complex optimization regulators for systems with limited computing resources.

This paper proposes an approach of cooperative control that is based on the concept of combining neural networks and the methodology of sliding mode control (SMC). The main purpose is to eliminate the chattering phenomenon. Next, the system performance can be improved by using the method of SMC. In the present approach, two parallel Neural Networks are utilized to realize a neuro-sliding mode control (NSMC), where the equivalent control and the corrective control are the outputs of neural network 1 and neural network 2, respectively. Based on expressions of the SMC, the weight adaptations of neural network can be determined. Furthermore, the gradient descent method is used to minimize the control force so that the chattering phenomenon can be eliminated. Finally, experimental results are given to show the effectiveness and feasibility of the approach.

Granular neural networks: A study of optimizing allocation of information granularity in input space

Accurate and non-contact measurement of volume percentages of oil-water fluids using microstrip sensors independent of the volume of sample using artificial neural network

Electromyogram (EMG) has been often used as a control signal of a prosthetic arm, which includes information on not only muscle force but operator's motor intention and mechanical impedance of joints. Most of previous researches, however, adopted the control methods of the prosthetic arms based on the EMG pattern discrimination and/or the force estimation from the EMG signals, and did not utilize any knowledge on tasks performed by amputees such as a grasping-an-object task and a spooning-soup task. In this paper, a new EMG pattern discrimination method is proposed using a statistically organized neural network and an event-driven task model. The neural network outputs aposterioriprobabilities of motions depending on the EMG signals. The task model is represented using a Petri net to describe the task dependent knowledge, which is used to modify the neural network's output. Experimental results show that the use of the task model significantly improves the accuracy of the EMG pattern discrimination.

In this paper, we propose a view invariant method for face discrimination using two kinds of three-layered neural networks. The first neural network learns informations of face obtained from different view point. Considering the human visual mechanism, two images from different views are fed into the first neural network. The second neural network integrates the outputs of the first neural network and recognizes a person. Our method has achieved a recognition rate of 100% on a database of 20 persons containing 150 images per person.

This paper presents a method of simple adaptive control (SAC) for nonlinear systems using neural networks. The control input is given by the sum of the output of a simple adaptive controller and the output of the neural network. The neural network is used to compensate the nonlinearity of plant dynamics that is not taken into consideration in the usual simple adaptive control. The role of the neural network is to construct a linearized model by minimizing the output error caused by nonlinearities in the control system.

This article uses neural networks in forecasting stock market prices. With their ability to discover patterns in nonlinear and chaotic systems, neural networks offer the ability to predict market directions more accurately than current techniques such as technical analysis, fundamental analysis, and regression compared with neural network performance. Proposed intelligent stock market prediction system is based on the quantitative and qualitative factors. Three feedforward neural models can be used to analyze these factors. Input data to the neural network proposed are quantitative factors. Input data to the neural network proposed for qualitative factors can be factors related to the political effect considered. Third neural network consists of decision integration in which input data will be the outputs of above-mentioned neural networks. This facilitates to make right decision whether stock market is influenced by quantitative or qualitative factors

A robust neural network (NN) output feedback scheme is proposed for the motion control of rigid robots. A dynamic NN observer is presented to estimate the joint speeds. The stability of a closed-loop system composed of a robot, a NN observer, and a NN controller is proven. The NN weights in both the observer and the controller are tuned online, with no off-line learning phase required. Most importantly, we can guarantee the boundness of the estimated velocities, the position tracking errors, and the NN weights. Also no exact knowledge of the robot dynamics is required so that the NN controller is model-free and so applicable to any type of rigid robot. When compared with adaptive-type controllers, we do not require persistent excitation conditions, linearity in the unknown system parameters, or the tedious computation of a regression matrix. Thus the new NN approach represents an improvement over adaptive techniques.

In this paper, an adaptive neural network (NN) output feedback control is investigated for incommensurate fractional-order permanent magnet synchronous motors under the condition of input saturation. First, a NN state observer is presented to obtain the ‘virtual estimate’ of angle speed, where the unknown function is approximated by the NN. Then, in order to solve the input saturation problem, an auxiliary system is developed under fractional-order framework. Next, the command filtered technology with an error compensation mechanism is used to handle the ‘explosion of complexity’ in backstepping and remove the filtering errors. In addition, the frequency distributed model is utilized such that the Lyapunov theory is available in the backstepping design and the system stability is demonstrated. Finally, numerical simulations confirm the availability of the proposed design.

A probabilistic approach combining a hidden Markov model and neural networks is implemented to identify different functional entities in nucleotide sequences. This approach augments the Hidden Markov model probability parameters by using the outputs of neural networks. It is designed to capture the compositional properties of complex genes and thereby achieves low error rates and high correlation coefficient measures. Initial experiments demonstrate that the predictive efficiency of the model is considerably higher than the existing models of gene finding.

In the Russian Federation, the agro-industrial complex is one of the leading sectors of the eco-nomy with a volume of domestic product of 4.5%. Russia owns 10 % of all arable land in the world. According to the data on the sown areas by crops in 2020, most of the agricultural area of Russia is occupied by wheat. The Russian Federation ranks third in the ranking of leading countries in the production of this type of grain crops, as well as leading positions in its export. Brown (leaf) and linear (stem) rust is the most harmful disease of grain crops. It is the reason for the sparseness of wheat crops and leads to a sharp decrease in yield. Therefore, one of the main tasks of farmers is to preserve the crop from diseases. The application of such areas of artificial intelligence as computer vision, machine learning and deep learning is able to cope with this task. These artificial intelligence technologies allow us to successfully solve applied problems of the agro-industrial complex using automated analysis of photographic materials. Aim. To consider the application of computer vision methods for the problem of classification of lesions of cultivated plants on the example of wheat. Materials and methods. The CGIAR Computer Vision for Crop Disease dataset for the crop disease recognition task is taken from the open source Kaggle. It is proposed to use an approach to the re-cognition of lesions of cultivated plants using the well-known neural network models ResNet50, DenseNet169, VGG16 and EfficientNet-B0. Neural network models receive images of wheat as in-put. The output of neural networks is the class of plant damage. To overcome the effect of overfit-ting neural networks, various regularization techniques are investigated. Results. The results of the classification quality, estimated by the software using the F1-score metric, which is the average harmonic between the Precision and Recall measures, are presented. Conclusion. As a result of the conducted research, it was found that the DenseNet model showed the best recognition accuracy us-ing a combination of transfer learning technology and DropOut and L2 regulation technologies to overcome the effect of retraining. The use of this approach allowed us to achieve a recognition ac-curacy of 91%.