Search

The HNN is often used as a human-like associative memory (AM). Nonetheless, HNNs suffered from limited noise tolerance and limited storage capacity. However, unfortunately, Hopfield neural networks suffered from limited tolerance to noise and limited storage capacity. This study proposes a novel approach for overcoming these limitations using the concept of multiple connections architecture in a Hopfield neural network with an energy function and Hamming distance (HD). In this architecture, a single connection among neurons of the network is employed to store a pattern, resulting in an etalon array, and the collection of etalon arrays corresponding to each neuronal connection forms a weight matrix (also called connection matrix). This approach of storing a single pattern on a single set of connections reduces the chances of being trapped in local minima and increases storage capacity because more patterns are stored in the network as the number of connections increases. The Lyapunov energy function and the Hamming distance concept are utilized to determine the testing pattern’s proximity to one of the stored patterns for perfect recall. Several experiments are conducted to compare the recall success rate and storage capacity of handwritten character images to those of existing methods. Experiments using multiple connections Hopfield neural network with hamming distance and Lyapunov energy function demonstrate promising results in comparison to existing methods. [ABSTRACT FROM AUTHOR]

Cataracts are lenticular opacities that can appear in different parts of the lens in the eye and are a leading cause of blindness globally. Accurate detection and early‐stage diagnosis can prevent the cataract and improve the quality of life for cataract patients. However, clinical cataract detection and grading require the expertise of trained eye specialists, which may impede everyone's early intervention due to the underlying expenses. This article proposed a computer aid diagnosis method for cataract detection, which also grades the severity of cataracts from fundus retinal images such as normal, mild, moderate, and severe. The proposed method uses a hybrid approach in which various pre‐trained convolutional neural networks (AlexNet, VGGNet, ResNet) with transfer learning are used to extract features. These feature vectors of each network individually and in the fused form are applied on the support vector machine classifiers for 4‐stage cataract classification. This architecture also takes advantage of ensemble learning by applying a majority voting scheme on the predictions of these SVM classifiers. The fundus cataract images are obtained from several open‐access datasets and arranged into 4‐classes with the assistance of an eye specialist. Since all the collected images are not suitable for diagnosis, an image quality selection module is included with this method to determine the quality of fundus images. The proposed method achieved 96.25% 4‐class classification accuracy. According to the experimental results, the proposed method is effective for cataract classification and outperforms conventional methods. [ABSTRACT FROM AUTHOR]