Forward layer-wise learning of convolutional neural networks through separation index maximizing.

This paper proposes a forward layer-wise learning algorithm for CNNs in classification problems. The algorithm utilizes the Separation Index (SI) as a supervised complexity measure to evaluate and train each layer in a forward manner. The proposed method explains that gradually increasing the SI through layers reduces the input data's uncertainties and disturbances, achieving a better feature space representation. Hence, by approximating the SI with a variant of local triplet loss at each layer, a gradient-based learning algorithm is suggested to maximize it. Inspired by the NGRAD (Neural Gradient Representation by Activity Differences) hypothesis, the proposed algorithm operates in a forward manner without explicit error information from the last layer. The algorithm's performance is evaluated on image classification tasks using VGG16, VGG19, AlexNet, and LeNet architectures with CIFAR-10, CIFAR-100, Raabin-WBC, and Fashion-MNIST datasets. Additionally, the experiments are applied to text classification tasks using the DBPedia and AG's News datasets. The results demonstrate that the proposed layer-wise learning algorithm outperforms state-of-the-art methods in accuracy and time complexity.

Real-time hand rubbing quality estimation using deep learning enhanced by separation index and feature-based confidence metric.

Hand hygiene plays a crucial role in healthcare environments which can cease infections and diseases from spreading. It is also regarded as the second most effective way to control the transmission of COVID-19. The World Health Organization (WHO) recommends a 12-step guideline for alcohol-based hand rubbing. Compliance with this guideline is vital in order to clean the hands thoroughly. Hence, an automated system can help to improve the quality of this procedure. In this study, a large-scale and diverse dataset for both real and fake hand rubbing motions is collected as the first stage of building a reliable hand hygiene system. In the next stage, various pre-trained networks were analyzed and compared using a swift version of the Separation Index (SI) method. The proposed Swift SI method facilitates choosing the best pre-trained network without fine-tuning them on the whole dataset. Accordingly, the Inception-ResNet architecture achieved the highest SI among Inception, ResNet, Xception, and MobileNet networks. Fine-tuning the Inception-ResNet model led to an accuracy of 98% on the test dataset, which is the highest score in the literature. Therefore, from the proposed approach, a lightweight version of this model with fewer layers but almost the same accuracy is produced and examined. In the final stage, a novel metric, called Feature-Based Confidence (FBC), is devised for estimating the confidence of models in prediction. The proposed confidence measure is able to profoundly differentiate models with similar accuracy and determine the superior one. Based on the metrics results, the Inception-ResNet model is about 2x slower but 5% more confident than its lightweight version. Putting all together, by addressing the real-time application concerns, a Deep Learning based method is offered to qualify the hand rubbing process. The model is also employed in a commercial machine, called DeepHARTS, to estimate the quality of the hand rubbing procedure in different organizations and healthcare environments.

A hand rubbing classification model based on image sequence enhanced by feature-based confidence metric.

Hand hygiene is critical for declining the spread of viruses and diseases. Over recent years, it has been globally known as one of the most effective ways against COVID-19 outbreak. The World Health Organization (WHO) has suggested a 12-step guideline for hand rubbing. Due to the importance of this guideline, several studies have been conducted to measure compliance with it using Computer Vision. However, almost all of them are based on processing single images as input, referred to as baseline models in this paper. This study proposes a sequence model in order to process sequences of consecutive images as input. The model is a mixture of Inception-ResNet architecture for spatial feature extraction and LSTM for detecting time-series information. After training the model on a comprehensive dataset, an accuracy of 98.99% was achieved on the test set. Compared to the best baseline models, the proposed sequence model is correspondingly about 1% and 4% better in terms of accuracy and confidence, though 3 times slower in inference time. Furthermore, this study demonstrates that the accuracy metric is not necessarily adequate to compare different models and optimize their hyperparameters. Accordingly, the Feature-Based Confidence Metric was utilized in order to provide a more pleasing comparison to discriminate the proposed sequence model with the best baseline model and optimize its hyperparameters.

Developing a carpet cloak operating for a wide range of incident angles using a deep neural network and PSO algorithm.

Designing invisibility cloaks has always been one of the most fascinating fields of research; in this regard, metasurface-based carpet cloaks have drawn researchers' attention due to their inherent tenuousness, resulting in a lower loss and easier fabrication. However, their performances are dependent on the incident angle of the coming wave; as a result, designing a carpet cloak capable of rendering objects under it invisible for a wide range of angles requires advanced methods. In this paper, using the Particle Swarm Optimization (PSO) algorithm, along with a trained neural network, a metasurface-based carpet cloak is developed capable to operate for a wide range of incident angles. The deep neural network is trained and used in order to accelerate the process of calculation of reflection phases provided by different unit cell designs. The resultant carpet cloak is numerically analyzed, and its response is presented and discussed. Both near-field and far-field results show that the designed carpet cloak operates very well for all incident angles in the range of 0 to 65 degrees.

Unsupervised Generative Adversarial Network for Plantar Pressure Image-to-Image Translation.

Analyzing human gait from plantar pressure is critical for human health. The majority of works focus on classifying the healthy plantar pattern from unhealthy ones. Different from previous works, we adopt a generative adversarial network to produce healthy plantar pressure image for individual patients. In this work, we do not have pairs of images for training thus we cast the problem as an unsupervised generative adversarial learning task. Our network benefits from multiple components: an encoder-decoder generator, a convolution-based discriminator, a convolution-based evaluation network, and a new term in the loss function to preserve the person's gait style. Our method achieves high performance (99.8%) on the CAD WALK databases which have patients with hallux valgus disease.

Control of a two-DOF parallel robot with unknown parameters using a novel robust adaptive approach.

Model-based methods lose their performance in confronting with model uncertainties and disturbances. Accordingly, some degrees of adaptation to the involved conditions are required. In this paper, a novel robust adaptive scheme is proposed which guarantees the simultaneous identification and control of a system in the presence of external disturbances. Thereafter, the suggested algorithm is implemented on a 2-DOf spherical parallel robot as a stabilizer device. By identifying unknown parameters of Jacobian matrix, the relative identification error is obtained as 0.0207. Applying external excitations to the base, the ratio of end-effector to base orientation is acquired as 0.091, demonstrating proper stabilization in comparison with other two well-known methods. The proposed structure also reveals a reliable performance in tracking desired paths for the end-effector Euler angles.

Prediction of breeding values for dairy cattle using artificial neural networks and neuro-fuzzy systems.

Developing machine learning and soft computing techniques has provided many opportunities for researchers to establish new analytical methods in different areas of science. The objective of this study is to investigate the potential of two types of intelligent learning methods, artificial neural networks and neuro-fuzzy systems, in order to estimate breeding values (EBV) of Iranian dairy cattle. Initially, the breeding values of lactating Holstein cows for milk and fat yield were estimated using conventional best linear unbiased prediction (BLUP) with an animal model. Once that was established, a multilayer perceptron was used to build ANN to predict breeding values from the performance data of selection candidates. Subsequently, fuzzy logic was used to form an NFS, a hybrid intelligent system that was implemented via a local linear model tree algorithm. For milk yield the correlations between EBV and EBV predicted by the ANN and NFS were 0.92 and 0.93, respectively. Corresponding correlations for fat yield were 0.93 and 0.93, respectively. Correlations between multitrait predictions of EBVs for milk and fat yield when predicted simultaneously by ANN were 0.93 and 0.93, respectively, whereas corresponding correlations with reference EBV for multitrait NFS were 0.94 and 0.95, respectively, for milk and fat production.