This paper presents the improved version of the classification system for supporting glaucoma diagnosis in ophthalmology. In this
paper we propose the new segmentation step based on the support vector clustering algorithm which enables better classification performance.
Prof. Małgorzata Kossut of the Nencki Institute of Experimental Biology talks about brain plasticity, the mechanisms of learning, and the mysteries of forgetfulness.
In recent years, deep learning and especially deep neural networks (DNN) have obtained amazing performance on a variety of problems, in particular in classification or pattern recognition. Among many kinds of DNNs, the convolutional neural networks (CNN) are most commonly used. However, due to their complexity, there are many problems related but not limited to optimizing network parameters, avoiding overfitting and ensuring good generalization abilities. Therefore, a number of methods have been proposed by the researchers to deal with these problems. In this paper, we present the results of applying different, recently developed methods to improve deep neural network training and operating. We decided to focus on the most popular CNN structures, namely on VGG based neural networks: VGG16, VGG11 and proposed by us VGG8. The tests were conducted on a real and very important problem of skin cancer detection. A publicly available dataset of skin lesions was used as a benchmark. We analyzed the influence of applying: dropout, batch normalization, model ensembling, and transfer learning. Moreover, the influence of the type of activation function was checked. In order to increase the objectivity of the results, each of the tested models was trained 6 times and their results were averaged. In addition, in order to mitigate the impact of the selection of learning, test and validation sets, k-fold validation was applied.
The paper presents a method for designing a neural speed controller with use of Reinforcement Learning method. The controlled object is an electric drive with a synchronous motor with permanent magnets, having a complex mechanical structure and changeable parameters. Several research cases of the control system with a neural controller are presented, focusing on the change of object parameters. Also, the influence of the system critic behaviour is researched, where the critic is a function of control error and energy cost. It ensures long term performance stability without the need of switching off the adaptation algorithm. Numerous simulation tests were carried out and confirmed on a real stand.
The paper proposes to apply an algorithm for predicting the minimum level of the state of charge (SoC) of stationary supercapacitor energy storage system operating in a DC traction substation, and for changing it over time. This is done to insure maximum energy recovery for trains while braking. The model of a supercapacitor energy storage system, its algorithms of operation and prediction of the minimum state of charge are described in detail; the main formulae, graphs and results of simulation are also provided. It is proposed to divide the SoC curve into equal periods of time during which the minimum states of charge remain constant. To predict the SoC level for the subsequent period, the learning algorithm based on the neural network could be used. Then, the minimum SoC could be based on two basic types of data: the first one is the time profile of the energy storage load during the previous period with the constant minimum SoC retained, while the second one relies on the trains’ locations and speed values in the previous period. It is proved that the use of variable minimum SoC ensures an increase of the energy volume recovered by approximately 10%. Optimum architecture and activation function of the neural network are also found.
The traditional self organizing map (SOM) is learned by Kohonen learning. The main disadvantage of this approach is in epoch based learning when the radius and rate of learning are decreasing functions of epoch index. The aim of study is to demonstrate advantages of diffusive learning in single epoch learning and other cases for both traditional and anomalous diffusion models. We also discuss the differences between traditional and anomalous learning in models and in quality of obtained SOM. The anomalous diffusion model leads to less accurate SOM which is in accordance to biological assumptions of normal diffusive processes in living nervous system. But the traditional Kohonen learning has been overperformed by novel diffusive learning approaches.
This paper proposes a comprehensive study on machine listening for localisation of snore sound excitation. Here we investigate the effects of varied frame sizes, and overlap of the analysed audio chunk for extracting low-level descriptors. In addition, we explore the performance of each kind of feature when it is fed into varied classifier models, including support vector machines, k-nearest neighbours, linear discriminant analysis, random forests, extreme learning machines, kernel-based extreme learning machines, multilayer perceptrons, and deep neural networks. Experimental results demonstrate that, wavelet packet transform energy can outperform most other features. A deep neural network trained with subband energy ratios reaches the highest performance achieving an unweighted average recall of 72.8% from four types for snoring.