Artificial neural networks (ANNs), such as the convolutional neural network (CNN) and long short-term memory (LSTM), have high complexity and contain large numbers of parameters. Memristor-based neural networks, which...
详细信息
Artificial neural networks (ANNs), such as the convolutional neural network (CNN) and long short-term memory (LSTM), have high complexity and contain large numbers of parameters. Memristor-based neural networks, which have the ability of in-memory and parallel computing, are therefore proposed to accelerate the operations of ANNs. In this paper, a memristor-based hardware realization of long short-term memory (LSTM) network with in situ training is presented. The designed memristor-based LSTM (MbLSTM) network is composed of memristor-based LSTM cell and memristor-based dense layer. Sigmoid and tanh (hyperbolic tangent) activation functions are approximately implemented through intentionally designing circuit parameters. A weight update scheme with row-parallel characteristic is put forward to update the conductance of memristors in crossbars. The highlights of MbLSTM include an effective hardware-based inference process and in situ training. The validity of MbLSTM is substantiated through classification tasks. The robustness of MbLSTM to conductance variations is also analyzed. (C) 2020 Elsevier Ltd. All rights reserved.
The sequence data processing, such as signal classification, is an important part of pattern recognition. Long short-term memory recurrent neural networks (LSTM-RNN) are widely applicable across the sequencedata proc...
详细信息
The sequence data processing, such as signal classification, is an important part of pattern recognition. Long short-term memory recurrent neural networks (LSTM-RNN) are widely applicable across the sequence data processing due to the ability to learn long-term dynamics while avoiding vanishing and exploding gradient problems. However, the high cost of LSTM-RNN models in computation is the main obstacle to deploy LSTM-RNN models on devices with limited resources. In this paper, we propose performance transferring of LSTM-RNN models based on knowledge distillation for LSTM-RNN model acceleration to overcome this obstacle. Firstly, we propose a paradigm for transferring the performance of LSTM-RNN models to lightweight convolutional neural network (CNN) models. Then, based on the paradigm, we define a novel loss that utilizes the prediction of an LSTM-RNN model to train a lightweight CNN model. Experiments results on two sequence data processing tasks, automatic modulation classification and text classification, show that the proposed paradigm is effective and the proposed loss makes CNN models with low time consumption and few parameters achieve higher accuracies and generate similar category distributions to LSTM-RNN models. Consequently, CNN models trained in the proposed method can be utilized to replace LSTM-RNN models for LSTM-RNN model acceleration.
In this paper, we provide a new spatial data generalization method which we applied in hand gesture recognition tasks. data gathering can be a tedious task when it comes to gesture recognition, especially dynamic gest...
详细信息
In this paper, we provide a new spatial data generalization method which we applied in hand gesture recognition tasks. data gathering can be a tedious task when it comes to gesture recognition, especially dynamic gestures. Nowadays, the standard solutions when lacking data still consist of either the expensive gathering of new data or the impractical employment of hand-crafted data augmentation algorithms. While these solutions may show improvement, they come with disadvantages. We believe that a better extrapolation of the limited data's common pattern, through an improved generalization, should first be considered. We, therefore, propose a dynamic generalization method that allows to capture and normalize in real-time the spatial evolution of the input. The latter procedure can be fully converted into a neural network processing layer which we call Evolution Normalization Layer. Experimental results on the SHREC2017 dataset showed that the addition of the proposed layer improved the prediction accuracy of a standard sequence-processing model while requiring 6 times fewer weights on average for a similar score. Furthermore, when trained on only 10% of the original training data, the standard model was able to reach a maximum accuracy of only 36.5% alone and 56.8% when applying a state-of-the-art processing method to the data, whereas the addition of our layer alone permitted to achieve a prediction accuracy of 81.5%.
The wide applications of 3-D sonar measurements are severely limited by factors such as water column interference, acoustic shadows, complex structures, and scattering noise. Outliers in 3-D sonar data are difficult t...
详细信息
The wide applications of 3-D sonar measurements are severely limited by factors such as water column interference, acoustic shadows, complex structures, and scattering noise. Outliers in 3-D sonar data are difficult to remove using traditional methods because the inference factors are different from other types of point cloud data. Therefore, this article presents a novel outlier filtering method by analyzing the sequential characteristics of the 3-D sonar data. First, the underwater point cloud is processed by super-voxel clustering method to decompose complex point cloud structures into several super-voxels with simple structures. Then we convert the point cloud data into subsequencedata according to the surveying principle of 3-D sonar scanning and super-voxel results. After that, an anomaly score calculation and anomaly region determination method based on the rectangular information granulation of subsequencedata is proposed. This method can capture the intrinsic changing characteristics of each subsequence and has a good recognition effect on the abnormal subsequence. Finally, an outlier detection method combining the Grubbs principle and the abnormal score is proposed and applied to the abnormal subsequences, which considers the distortion not only in the vertical direction but also in the horizontal direction. The experimental results show that the proposed comprehensive filtering method has good accuracy for both horizontal and vertical point cloud data. The average overall accuracy of the test results is 99.1%, and the average kappa coefficient is 0.88, which can be effectively applied to the 3-D sonar point cloud data filtering processing in complex underwater areas.
Directed laboratory evolution is a common technique to obtain an evolved bacteria strain with a desired phenotype. This technique is especially useful as a supplement to rational engineering for complex phenotypes suc...
详细信息
Directed laboratory evolution is a common technique to obtain an evolved bacteria strain with a desired phenotype. This technique is especially useful as a supplement to rational engineering for complex phenotypes such as increased biocatalyst tolerance to toxic compounds. However, reverse engineering efforts are required in order to identify the mutations that occurred, including single nucleotide polymorphisms (SNPs), insertions/deletions (indels), duplications, and rearrangements. In this protocol, we describe the steps to (1) obtain and sequence the genomic DNA, (2) process and analyze the genomic DNA sequencedata, and (3) verify the mutations by Sanger resequencing. less
暂无评论