Search

Skin cancer is a huge issue which gets neglected very often. Sometimes the human eye is unable to precisely detect diseases from imaging data, in cases of doctor’s manual inspection. In this age, we see the rise of use of deep learning methods in our daily life problem solving. Therefore, we develop an automated computerised system for detecting skin diseases using deep neural network algorithms. In the proposed model, we have used several neural network algorithms and analyse their performances to detect five major skin diseases and Figure out the best performing algorithm in terms of accuracy. CNN and by using Keras Sequential API, we have structured a new model to gainan accuracy of around 80%. Later, for comparison and also to increase accuracy we have used architectures that use pre-trained data. These transfer learning model includes VGG11, RESNET50 and DENSENET121. Among the algorithms used in the proposed models, resnet architecture achieve highest accuracy of 90%.

Recent studies have demonstrated the importance of neural networks in medical image processing and analysis. However, their great efficiency in segmentation tasks is highly dependent on the amount of training data. When these networks are used on small datasets, the process of data augmentation can be very significant. We propose a convolutional neural network approach for the whole heart segmentation which is based upon the 3D U-Net architecture and incorporates principle component analysis as an additional data augmentation technique. The network is trained end-to-end i.e. no pre-trained network is required. Evaluation of the proposed approach is performed on 20 3D CT images from MICCAI 2017 Multi-Modality Whole Heart Segmentation Challenge dataset, divided into 15 training and 5 validation images. Final segmentation results show a high Dice coefficient overlap to ground truth, indicating that the proposed approach is competitive to state-of-the-art. Additionally, we provide the discussion of the influence of different learning rates on the final segmentation results.

Human detection in a scene is very useful in applications where the control of people flow or surveillance is required. Herein we propose a solution for person detection by counting pairs of legs using a simulation of a LIDAR from a Kinect laser scanner, and a particular instance of YOLO neural net architecture. We employed both geometrical and machine learning approaches for 2D and 3D images. We managed to achieve real-time detection of multiple people from 2D and 3D information. Our miss-rate for 3D images was as low as 3.2% and a false positive per image of 0.06 at 78 FPS.

Surface blemishes of various apple varieties were analyzed by their reflectance characteristics between 460 and 1130 nm. Normalized reflectance data were collected at 10 nm increments with liquid crystal tunable filters. Data were utilized as input values for various pattern recognition models specifically multi-layer back propagation, unimodal Gaussian, K-nearest neighbor and nearest cluster algorithms. Partitioning data into 50:50 training and test sets, correct classification in separating unflawed versus blemished areas ranged from 62 to 96% (Year I) and from 73 to 85% (Year II). The algorithm which yielded the highest correct classification was the multi-layer back propagation but minor variation was found for number of hidden nodes or neural net architecture.

We present a theoretical analysis of a version of the LAPART adaptive inferencing neural network. Our main result is a proof that the new architecture, called LAPART 2, converges in two passes through a fixed training set of inputs. We also prove that it does not suffer from template proliferation. For comparison, Georgiopoulos et al. (1994) have proved the upper bound n-1 on the number of passes required for convergence for the ARTMAP architecture, where n is the size of the binary pattern input space. If the ARTMAP result is regarded as an n-pass, or finite-pass, convergence result, ours is then a two-pass, or fixed-pass, convergence result. Our results have added significance in that they apply to set-valued mappings, as opposed to the usual supervised learning model of affixing labels to classes.

In this article we discuss the flow of design for neural network hardware and go deeper into the design constraints and implementation possibilities. The performance measures and problems of different measurements are also discussed. It is noted that performance is one comparison criteria, but there are also many others, some of which are also discussed. In order to anchor the discussion to real life, the article includes a case study of our TUTNC neurocomputer. In addition, examples of commercial neural computing systems and their world wide web pages are given.

Results are reported on the task of recognizing handwritten digits without any advanced pre-processing. The result are obtained using a RAM-based neural network, making use of small receptive fields. Furthermore, a technique that introduces negative weights into the RAM net is reported. The results obtained on the task of recognizing handwritten digits is comparable with the best performances reported in the literature.

Recently, circuit analysis and optimization featuring neural-network models have been proposed, reducing the computational time during optimization while keeping the accuracy of physics-based models. We present a novel approach for fast training of such neural-network models based on the sparse matrix concept. The new training technique does not require any structure change in neural networks, but makes use of the inherent nature of neural networks that for each pattern some neuron activations are close to zero, and hence, have no effect on network outputs and weights update. Much of the computation effort is saved over standard training techniques, while achieving the same accuracy. FET device and VLSI interconnect modeling examples verified the proposed technique.

A novel method for the detection of faults in DC and AC motors using Kohonen Networks (Self Organized Maps) is presented in this paper. The advantage of this technique is that it only requires input samples to perform the training of the network, a difference from other Neural Network architectures that need both inputs and outputs to perform the training. This technique generates fault maps based only on the inputs that are received from the motor under test. The maps can be used to clearly identify different types of faults in DC and AC motors since similar faults generate the same type of map. The main advantages of this technique are that it can be used to test motors in real time and on-site, without having to disconnect the motor for testing. The technique can be applied for testing motors used in production lines without having to stop operation for testing.

The suitability of partial least squares (PLS) and neural nets for the identification of mixtures of butter fat with foreign fat is compared. While neural nets are most suitable for classification, quantitative results are obtained by PLS. Butter fats of various European countries have been analyzed by GC. Fifty-six samples were used as the calibration set to build the PLS model and prepare the neural net, respectiveley. For successful modelling a 11-factor PLS model was sufficient. The neural net architecture chosen is a 17 × 1 preceptron. Both data evaluation techniques have been validated with 28 samples not included in the calibration set. For PLS the results indicate a detection limit of 1–2% foreign fat in butter fat. The neural net classified 20 samples correctly, but eight samples could not be classified at all.

Several continuous-time and discrete-time recurrent neural network models have been developed and applied to various engineering problems. One of the difficulties encountered in the application of recurrent networks is the derivation of efficient learning algorithms that also guarantee the stability of the overall system. This paper studies the approximation and learning properties of one class of recurrent networks, known as high-order neural networks; and applies these architectures to the identification of dynamical systems. In recurrent high-order neural networks, the dynamic components are distributed throughout the network in the form of dynamic neurons. It is shown that if enough high-order connections are allowed then this network is capable of approximating arbitrary dynamical systems. Identification schemes based on high-order network architectures are designed and analyzed. >

To solve the fault forecast problem caused by small samples, a model based on grey neural network GNNM (1,1) is presented. The quadratic momentum is introduced on the base of GNNM (1,1) model and the convergence of the model is proved. The experiment results show that the GNNM (1,1) model based on quadratic momentum has a better prediction results than traditional GNNM (1,1) model.

A novel bio-inspired computational high-precision dental milling system is proposed in this interdisciplinar research. The system applies several bio-inspired models, based on unsupervised learning, that analyse and identify the most relevant features of high-precision dental-milling data sets and their internal structures. Finally, a supervised neural architecture and certain identification techniques are applied, in order to model and to optimize the high-precision process. This is done by empirically testing the model using a real data set taken from a dynamic high-precision machining centre with five axes.

The design of a multi-chip architecture can be an appealing solution to implement very large Artificial Neural Networks systems. The size of these systems often can be a only way to deeply and effectively investigate on innovative, “bioinspired”, computational paradigms. In this paper, the authors present the architecture of a system with two analogue chips well suited to carry out a large Integrate and Fire Neural Network electronic system. In particular, the chips have been designed by using a dedicated technique to reduce the I/O analogue pins.

This paper researchs the approaches of data repairing for expendable current profiler. The self-learning algorithm of the network architecture based on generalization ability has been developed. Proving experiment using the algorithm has been done, which confirms the algorithm can increase the precise of data gained from the expendable current profiler. Keywords—data repairing; current profiler; neural network; generalizaition

In this paper the time-delay neural networks (TDNNs) have been implemented in detecting the damage in bridge structure using vibration signature analysis. A simulation study has been carried out for the incomplete measurement data. It has been observed that TDNNs have performed better than traditional neural networks in this application and the arithmetic of the TDNNs is simple.

A review of the existing situation in a computer system's security zone is made. Classification of security threats and possible attacks on them carried out. The advisability of using neural networks for detecting network attacks is substantiated The Back-Propagation Neural Network for solution of the problem put by is selected. Attacks in the network traffic were correctly identified.

Artificial neural networks are used in various applications and research areas. Mathematically inspired approaches use these types of networks to solve complex classification or function approximation tasks whereas biologically motivated models attempt to adapt desired properties from biology such as robustness or fault tolerance to technical systems and architectures. Therefore, a great variety of different models have been proposed in literature which can be separated in time-dependent and time-independent models. To verify these models and to accelerate simulations prototypes are often implemented in integrated circuits using digital or analog designs. In this work, a simple reconfigurable neural hardware structure (SIRENS) is introduced which is capable to represent several different models of neurons, time-independent and time-dependent models as well. Therefore, this system can be used for several applications (classification or simulation) and purposes (acceleration or operation). The underlying mathematical principles are presented and, furthermore, design considerations are given in this paper.

We present a new neural architecture which imbeds dynamic programming solutions to solve optimal target-intercept problems. They provide feedback guidance solutions, which are optimal with any initial conditions and time-to-go, for a 2D scenario. The method discussed in this study determines an optimal control law for a system by successively adapting two networks - an action and a critic network. This method determines the control law for an entire range of initial conditions; it simultaneously determines and adapts the neural networks to the optimal control policy for both linear and nonlinear systems. In addition, it is important to know that the form of control does not need to be known in order to use this method.

In this paper we outline some of the changes needed to implement multilayer feed-forward neural networks using the demonstrator hardware which was based on around an array of vertical cavity surface emitting lasers. Network simulations show that the neural network demonstrator hardware can be used to implement two different classes of feed-forward network, the multilayer perceptron (MLP) and radial basis function (RBF) networks. In both cases, the actual training of the networks is performed offline using hardware simulations and the weighted interconnections between neurons are fixed before application to the optoelectronic hardware.

Among many neural network architectures that exist in the literature, the recurrent neural networks (RNN's) are of special interest due to their ability to deal with spatial temporal problems. However, in an earlier published paper, the authors shown that RNN's have poor performance in terms of phoneme accuracy when applied to the specific problem of converting text streams into their phonetic transcriptions. This is due to the fact that RNN's contains a weak left side context dependence between letters and the right side context dependence is not included. In this paper, we study the behavior of RNN that includes the context information between adjacent letters at the input. The results in terms of phoneme accuracy, for the RNN with both side input context dependence, multilayer perception and RNN, in the context of text-to-phoneme mapping, are shown.

The application of two kinds of autoassociative NN's is discussed. The applications concern observation of the marine environment. The first kind of autoassociative NN has physical interpretable neurons in the bottleneck layer and is used for scope check in the retrieval of concentrations of water constituents. The second kind is a standard autoassociative NN which we propose to use in the monitoring of the environment. An example of such a usage is given.

This paper presents how to find an architecture for very large scale lossless neural nets, which can be used as Haar-Walsh spectrum analyzers. This analysis relies on the orthogonality of weight matrices W, where W could be Hurwitz-Radon matrices. The unique feature of these nets is the possibility to treat them either as algorithms or as Hamiltonian physical objects (Haar-Walsh Signal Processors).

VLSI architectures for neural networks are presented. Neural networks have wide-ranging applications in classification, control, and optimization. With the need for real-time performance, VLSI neural networks have gained significant attention. Digital, analog, and mixed-mode designs are used for this application. Modular and reconfigurable designs are necessary so that various neural network models can be easily configured. >

This paper presents and analyses distinct learning strategies which have been used by GSN/sup f/ architectures. Sharing the common feature of being one-shot learning, these strategies achieve different performances as key parameters are changed. These algorithms are evaluated against each other by taking into account the training time, saturation rates, learning conflict rates and recognition performance.

The neural net model TOM (temporal organization map) that we present in the paper is a new connectionist approach whose time representation is different from the one in classical temporal connectionist models. The architecture is neurobiologically inspired and is dedicated to sensory problems involving a temporal dimension. The basic idea of the TOM model is the propagation of an activity throughout the network whose elements are organized according to a map architecture. This propagation leads to a triggering of a sequence detection. We have applied this new kind of architecture to a spoken digit recognition problem. The results draw near to the results of the best hidden Markov model (HMM) techniques. The interest of such an architecture is its genericity and the possibility to merge several data flows in order to improve the classical performances of neural nets.

Over the last decade neural networks have emerged as a versatile computing technology, offering efficient solutions in many engineering, industrial and business applications, including decision support systems and medicine. This paper addresses the basic concepts of neurocomputing and discusses applications relevant to the aerospace industry. The advantages and drawbacks of neurocomputing are discussed as compared to conventional approaches to engineering solutions. Also, current needs and potential developments in the field are reviewed.

An analog VLSI neural network integrated circuit is presented. It consist of a feedforward multi layer perceptron (MLP) network with 64 inputs, 64 hidden neurons and 10 outputs. The computational cells have been designed by using the current mode approach and weak inversion biased MOS transistors to reduce the occupied area and power consumption. The processing delay is less than 2 /spl mu/s and the total average power consumption is around 200 mW. This is equivalent to a computational power of about 2.5/spl times/10/sup 9/ connections per second. The chip can be employed in a chip-in-the-loop neural architecture.

This paper introduces an analytic method to determine the sensitivity to random parameter variations of analog VLSI neural network architectures for linear image filtering. The authors compare the robustness of several different circuit architectures for low pass filtering. This method can also determine which components within a particular architecture should specified the most precisely. >

In this contribution neuronic cellular automata, a particular subcase of DTCNN, are studied from the point of view of the influence of the distribution of their connection weights on their dynamic behavior and on their spatial correlation length. >

We define layered networks with horizontal connections as networks having units that receive inputs from the lower layer and also from the previous units of the same layer. We show that architecture with horizontal connections does not require as many units in the hidden layers as the plain layered architecture in order to approximate a function. >

The paper discusses the detection of weaknesses and defects on photolithographic masks by cellular neural networks. The detection by cellular neural networks is performed with the advantages of their massive parallel architecture. First a survey is given of actual methods for the detections of weaknesses and defects. Then the relations between the structures of the mask layouts and the real structures of the masks are defined by local design rules. These local design rules can also indirectly be used to detect most weaknesses and defects. After that, the design of the operators for the detection of weaknesses and defects are executed on the basis of the local design rules, using the method of Galias that is practicable by cellular neural networks. Then some examples of weakness and defect detections on real mask images by cellular neural networks are presented. Finally the results and future aims are discussed. >

We present significant optimization of the so-called combinatorial neural model (CNM). CNM is a hybrid (neural/symbolic) model that has been used in areas such as expert system development and data mining. The paper first explains the CNM architecture and then presents CNM optimization together with empiric results. The most important optimization aims at taming combinatorial explosion, which is the main problem inherent to this model.

In neural networks design, some parameters must be adequately set for an efficient performance to be achieved. The setting of these parameters is not a trivial task since different applications may require different values. The "trial-and-error" or traditional engineering approaches for this task do not guarantee that an optimal set of parameters is found. Evolutionary approaches have been recently proposed to overcome these problems. Genetic algorithms have been used as a heuristic search technique to find adequate neural architectures. This technique is based on natural selection and genetic mechanisms. It tries to produce better individuals from an initial set of individuals, based on a given criteria. This paper presents some results achieved by using this technique to search optimal neural architectures to solve real world credit analysis problems. In this paper the searches have been restricted to multilayer perceptron (MLP) networks which are feedforward fully-connected strictly-layered architectures.

A model based on the application of the Dynamic Link Architecture (DLA) is presented for the off-line recognition of Chinese characters. It is a revised DLA model employing 4-vector dynamic link assignment instead of the original 3-vector links. A sample set of Chinese characters was used to test the performance of the model under various transformations including translation, reflection, rotation, dilation and distortion. Challenging results are obtained. An improvement of 40% in recognition rate was attained by using the revised DLA model for Chinese character recognition. On the other hand, for the testing of invariant properties, an overall correct recognition rate of 85% was obtained under various transformations.

This paper presents the random-pulse machine concept and shows how it can be used for the modular design of artificial neural networks. Random-pulse machines deal with analog variables represented by the mean rate of random-pulse streams and use simple digital technology to perform arithmetic and logic operations. As an application example, a NN is proposed for modeling of the activated sludge wastewater treatment plants.

The design options of superresolution using a neural network are discussed. Four specific neural-network architectures using different preprocessors are described, and their performances are compared. The correlation transformer is found to be the preprocessor that provides the best performance and the simplest implementation. The correlation transformer converts N complex inputs derived from a phased-array antenna into N(N+1)/2 complex outputs that are applied to the neural-network.

The adjusting neural network (AJNN) we (1995) proposed previously has the capability for parameter tuning of a control model, namely it can perform parameter tuning accurately with small tuning numbers. However, when parameter errors are relatively large, its tuning capability may occasionally deteriorate, which leads to an increase of tuning numbers. In this paper, we discuss two ways of overcoming this weakness of the AJNN. We propose a new learning algorithm for the AJNN and develop the AJNN architecture. We simulate the effectiveness of both approaches and compare these results with results from our previous AJNN using the problem of temperature control for a reheating furnace plant.

Bounds on the number of training examples needed to guarantee a certain level of generalization performance in the ARTMAP architecture are derived. Conditions are derived under which ARTMAP can achieve a specific level of performance assuming any unknown, but fixed, probability distribution on the training data.

The application of artificial neural networks (ANN) in real-time embedded systems demands high performance computers. Miniaturized massively parallel architectures are suitable computation platforms for this task. An important question which arises is how to establish an effective mapping from ANN algorithms to hardware. In this paper, we demonstrate how an effective mapping can be achieved with our programming environment in close combination with an optimized architecture design targeted for neuro-computing. >

A neural network that learns to play the board game of five-in-a-row is presented. The basic idea of the approach is to let an appropriately designed network play a series of games against an opponent and use a reinforcement learning algorithm to train the network to evaluate the non-occupied board positions by rewarding good moves and penalizing bad moves. The performance of the proposed network is demonstrated by presenting experimental results.

A hybrid architecture for Hopfield artificial neural networks and its GaAs implementation are described in this paper. This approach merges parallel summation and multiplication and serial neuron state updating. For a N/spl times/N network, only N multipliers and one neuron activation function are needed.

A controller for an agile, high-subsonic autonomous flight vehicle, incorporating neural network and genetic algorithm techniques, is presented. Simulated flight results for nominal and off-nominal vehicle configurations are reported. The results show that an inverse dynamic model neural network can offer better tracking performance and greater robustness than a conventional linear controller. However, the genetic algorithm technique employed here was found to offer no significant improvement in controller performance. >

Describes an evolved artificial neural network that is evolved by the particle swarm optimisation (PSO) algorithm. Compared with previous evolved ANNs, both the architecture and weights of this ANN are evolved by PSO, it means that the network architecture is adaptively adjusted, then the PSO algorithm is employed to evolve the nodes of ANNs with a given architecture. This process is repeated until the best network is accepted or the maximum number of generations has been reached. Some techniques, such as partial training algorithm (PT) and evolving added nodes (EAN), are used to maintain a closer behavioural link between the parents and their offspring, which will improve the efficiency of evolving ANNs. An ANN evolved is used in modelling a product quality estimator for a fractionator of the hydrocracking unit in the oil refining industry. The results show that the evolved ANN has good accuracy and generalisation ability.

Conference 2002 IEEE Intenacional Conference on Systems, Man and Cybernetics, Túnez, 6-9 Oct. 2002 Many methods to codify artificial neural networks have been developed to avoid the defects of direct encoding schema, improving the search into the solution's space. A method to evaluate how the search space is covered and how movement along the search process applying genetic operators is needed in order to evaluate the different encoding strategies for a kind of artificial neural networks, feedforward neural networks. A first step of this method is considered with two encoding strategies, a direct encoding method and an indirect encoding scheme based on cellular automata. Publicado

[Abstract] This article implements an artificial neural network to find, through computer simulation, the excitations of a square planar array. The array is composed of 52 uniformly spaced subarrays, and has a quasi- in its radiation diagram. This simulation model includes the reduction of any signal interference in the shaped radiating zone after its position has been determined.