Search

Recently, many studies have been carried out on model compression to handle the high computational cost and high memory footprint brought by the implementation of deep neural networks. In this paper, model compression of convolutional neural networks is constructed as a multiobjective optimization problem with two conflicting objectives, reducing the model size and improving the performance. A novel structured pruning method called Conventional-based and Evolutionary Approaches Guided Multiobjective Pruning (CEA-MOP) is proposed to address this problem, where the power of conventional pruning methods is effectively exploited for the evolutionary process. A delicate balance in pruning rate and model accuracy has been automated achieved by a multiobjective optimization evolutionary model. First, an ensemble framework integrates pruning metrics to establish a codebook for further evolutionary operations. Then, an efficient coding method is developed to shorten the length of chromosome, thus ensuring its superior scalability. Finally, sensitivity analysis is automatically carried out to determine the upper bound of pruning rate for each layer. Notably, on CIFAR-10, CEA-MOP reduces more than 50% FLOPs on ResNet-110 and improves the relative accuracy. Moreover, on ImageNet, CEA-MOP reduces more than 50% FLOPs on ResNet-101 with negligible top-1 accuracy drop. [ABSTRACT FROM AUTHOR]

Synchronous parameters algorithms with data parallelism have been successfully utilized to accelerate the distributed training of deep neural networks (DNNs). However, a prevalent shortcoming of the synchronous methods is computation waste resulted from the mutual waiting among the computational workers with different performance and the communication delays at each synchronization. To alleviate this drawback, we propose a novel method, free local stochastic gradient descent (FLSGD) with parallel synchronization, to eliminate the waiting and communication overhead. Specifically, the process of distributed DNN training is firstly modeled as a pipeline which assembly consists of three components: dataset partition, local SGD, and parameter updating. Then, a novel adaptive batch size and dataset partition method based on the computational performance of the node is employed to eliminate the waiting time by keeping the load balance of the distributed DNN training. The local SGD and the parameter updating including gradients synchronization are parallelized to eliminate the communication cost by one-step gradient delaying, and the stale problem is remedied by an appropriate approximation. To our best knowledge, this is the first work focusing on decreasing both distributed training load balancing and communication overhead Extensive experiments are conducted with four state-of-the-art DNN models on two image classification datasets (i.e., CIFAR10 and CIFAR100) to demonstrate that the effectiveness of FLSGD outperforms the synchronous methods. [ABSTRACT FROM AUTHOR]

The main challenge for most image–text tasks, such as zero-shot, is the way to measure the semantic similarity between visual and textual feature vectors. The common solution is to map the image feature vectors and text feature vectors into the Hilbert space and then rank the similarity by the inner product between feature vectors. In this paper, we learn the feature representation of images and their sentence descriptions by different deep neural networks to learn about the inner-modal correspondences between visual and language data. We then use a joint embedding structure based on angle calculation for measuring the semantic similarity between visual and textual features. In the proposed method, a constant factor b keeps the similarities of positive samples and negative samples at a certain distance. Since the proposed cosine similarity method involves both normalization and vectors computation, we also develop the learning algorithm on neural networks for expressing the semantic features of texts and images. We applied the angle-based method to the challenging Caltech-UCSD Birds and the Oxford-102 Flowers datasets. The experiments demonstrate good performances on both recognition and retrieval tasks. [ABSTRACT FROM AUTHOR]

Finding shortest paths is an important problem in transportation and communication networks. This paper develops a Pulse-Coupled Neural Network (PCNN) model to efficiently compute a single-pair shortest path. Unlike most of the existing PCNN models, the proposed model is endowed with a special mechanism, called on-forward/off-backward ; if a neuron fires, its neighboring neurons in a certain forward region will be excited, whereas the neurons in a backward region will be inhibited. As a result, the model can produce a restricted autowave that propagates at different speeds corresponding to different directions, which is different from the completely nondeterministic PCNN models. Compared with some traditional methods, the proposed PCNN model significantly reduces the computational cost of searching for the shortest path. Experimental results further confirmed the efficiency and effectiveness of the proposed model. [ABSTRACT FROM AUTHOR]

Data-driven method has been widely used in Fluid Catalytic Cracking (FCC) process modeling. However, due to the complexity of chemical process both in time and spatial domain, how to reflect the time and spatial characteristics of FCC units and build corresponding model is important to construct a better model for the gasoline yield prediction. In this paper, a special neural network structure was developed to deal with the input variables with different time scales considering the collection characteristics of various variables, as well as the time continuity of large-scale process manufacturing units, LSTMs with different time scales are stacked to extract temporal and spatial features to help capture the relationship between influencing factors and product yield. The characteristics of FCC process are also fully reflected in data processing and building model. It is demonstrated from the conclusions that the new model developed in this paper performs better than the traditional LSTM networks, which will be of great help to the intelligent upgrading of the FCC process. • A special network structure is built to deal with the difference of data recording frequency. • A new neural network model is proposed to predict the product yield of FCCU,. • Analyzing the relationship between the product yield and the features extracted by our new neural network. [ABSTRACT FROM AUTHOR]

Recent research on hand detection and gesture recognition has attracted increasing interest due to its broad range of potential applications, such as human-computer interaction, sign language recognition, hand action analysis, driver hand behavior monitoring, and virtual reality. In recent years, several approaches have been proposed with the aim of developing a robust algorithm which functions in complex and cluttered environments. Although several researchers have addressed this challenging problem, a robust system is still elusive. Therefore, we propose a deep learning-based architecture to jointly detect and classify hand gestures. In the proposed architecture, the whole image is passed through a one-stage dense object detector to extract hand regions, which, in turn, pass through a lightweight convolutional neural network (CNN) for hand gesture recognition. To evaluate our approach, we conducted extensive experiments on four publicly available datasets for hand detection, including the Oxford, 5-signers, EgoHands, and Indian classical dance (ICD) datasets, along with two hand gesture datasets with different gesture vocabularies for hand gesture recognition, namely, the LaRED and TinyHands datasets. Here, experimental results demonstrate that the proposed architecture is efficient and robust. In addition, it outperforms other approaches in both the hand detection and gesture classification tasks. [ABSTRACT FROM AUTHOR]

Deep reinforcement learning (DRL) has shown promising results in solving robotic control and decision tasks, which can learn the high-dimensional state and action information well. Despite their successes, conventional neural-based DRL models are criticized for low energy efficiency, making them laborious to be widely applied in low-power electronics. With more biologically plausible plasticity principles, spiking neural networks (SNNs) are now considered an energy-efficient and robust alternative. The most existing dynamics and learning paradigms for spiking neurons with a common Leaky Integrate-and-Fire (LIF) neuron model often result in relatively low efficiency and poor robustness. To address these limitations, we propose a multi-attribute dynamic attenuation learning improved spiking actor network (MADA-SAN) for reinforcement learning to achieve effective decision-making. The resistance, membrane voltage and membrane current of spiking neurons are updated from a fixed value into dynamic attenuation. By enhancing the temporal relation dependencies in neurons, this model can learn the spatio-temporal relevance of complex continuous information well. Extensive experimental results show MADA-SAN performs better than its counterpart deep actor network on six continuous control tasks from OpenAI gym. Besides, we further validated the proposed MADA-LIF can achieve comparable performance with other state-of-the-art algorithms on MNIST and DVS-gesture recognition tasks. [ABSTRACT FROM AUTHOR]

Abstract: The eigenvector associated with the smallest eigenvalue of the autocorrelation matrix of input signals is called minor component. Minor component analysis (MCA) is a statistical approach for extracting minor component from input signals and has been applied in many fields of signal processing and data analysis. In this letter, we propose a neural networks learning algorithm for estimating adaptively minor component from input signals. Dynamics of the proposed algorithm are analyzed via a deterministic discrete time (DDT) method. Some sufficient conditions are obtained to guarantee convergence of the proposed algorithm. [Copyright &y& Elsevier]

• The proximal femur has considerable value in estimating sex. • Deep learning can be used to create an automated sex estimation model based on femur radiographic features. • The model proposed in this paper achieves competitive performance compared with existing models. To develop a fully automated deep learning pipeline using digital radiographs to detect the proximal femur region for accurate automated sex estimation. Radiograph predictive features from 2122 Chinese Han clinical pelvic with ages ranging from 18 to 26 years were collected retrospectively to train and test the sex prediction model using deep machine learning's convolutional neural networks (CNN). Model performance was assessed using a Chinese Han population with 361 samples and a white population with 50 samples. The average accuracy of the sex estimation of the two test datasets was determined. For the Chinese Han population test dataset, the sex estimation accuracy was 94.6% (males: 93.9% and females: 94.7%). For the white population samples, the accuracy of sex estimation was 82.9% (males: 80.9% and females: 88.6%). The accuracy of CNN tested in the Chinese population was significantly higher than that tested in the White population (p < 0.001) The model based on convolutional neural networks has an accuracy similar to that of current state-of-the-art mathematical functions using manually extracted features for the Chinese Han population samples, proving to be a reliable choice for the human sex estimation. [ABSTRACT FROM AUTHOR]

• Deep adversarial learning technique is introduced to synthesize the face images from the skull images for automated craniofacial reconstruction. • Artificial neural network architecture and one-to-many inference approach are designed. • Through comprehensive automated face recognition tests on five deep face recognition algorithms for the recreated face, we demonstrate the effectiveness of the proposed approach. Craniofacial reconstruction is applied to identify human remains in the absence of determination data (e.g., fingerprinting, dental records, radiological materials, or DNA), by predicting the likeness of the unidentified remains based on the internal relationship between the skull and face. Conventional 3D methods are usually based on statistical models with poor capacity, which limit the description of such complex relationship. Moreover, the required high-quality data are difficult to collect. In this study, we present a novel craniofacial reconstruction paradigm that synthesize craniofacial images from 2D computed tomography scan of skull data. The key idea is to recast craniofacial reconstruction as an image translation task, with the goal of generating corresponding craniofacial images from 2D skull images. To this end, we design an automatic skull-to-face transformation system based on deep generative adversarial nets. The system was trained on 4551 paired skull-face images obtained from 1780 CT head scans of the Han Chinese population. To the best of our knowledge, this is the only database of this magnitude in the literature. Finally, to accurately evaluate the performance of the model, a face recognition task employing five existing deep learning algorithms, —FaceNet, —SphereFace, —CosFace, —ArcFace, and —MagFace, was tested on 102 reconstruction cases in a face pool composed of 1744 CT-scan face images. The experimental results demonstrate that the proposed method can be used as an effective forensic tool. [ABSTRACT FROM AUTHOR]