Stochastic gradient descent (SGD) is a popular and efficient method with wide applications in training deep neural nets other nonconvex models. While the behavior of SGD well understood convex learning setting, existing theoretical results for applied to objective functions are far from mature. For example, require impose nontrivial assumption on uniform boundedness gradients all iterates encountered process, which hard verify practical implementations. In this article, we establish rigorous...

Layer-wise magnitude-based pruning (LMP) is a very popular method for deep neural network (DNN) compression. However, tuning the layer-specific thresholds difficult task, since space of threshold candidates exponentially large and evaluation expensive. Previous methods are mainly by hand require expertise. In this paper, we propose an automatic approach based on optimization, named OLMP. The idea to transform problem into constrained optimization (i.e., minimizing size pruned model subject...

Deep neural networks (DNNs) have achieved great success, but the applications to mobile devices are limited due their huge model size and low inference speed. Much effort thus has been devoted pruning DNNs. Layer-wise neuron methods shown effectiveness, which minimize reconstruction error of linear response with a number neurons in each single layer pruning. In this paper, we propose new layer-wise approach by minimizing nonlinear units, might be more reasonable since before after activation...

Neural network pruning is a popular approach to reducing the computational complexity of deep neural networks. In recent years, as growing evidence shows that conventional methods employ inappropriate proxy metrics, and new types hardware become increasingly available, hardware-aware incorporates characteristics in loop has gained attention. Both accuracy efficiency (latency, memory consumption, etc.) are critical objectives success pruning, but conflict between multiple makes it impossible...

A recurrent neural network (RNN) has shown powerful performance in tackling various natural language processing (NLP) tasks, resulting numerous models containing both RNN neurons and feedforward neurons. On the other hand, deep structure of heavily restricted its implementation on mobile devices, where quite a few applications involve NLP tasks. Magnitude-based pruning (MP) is promising way to address such challenge. However, existing MP methods are mostly designed for networks that do not...

The subset selection problem that selects a few items from ground set arises in many applications such as maximum coverage, influence maximization, sparse regression, etc. recently proposed POSS algorithm is powerful approximation solver for this problem. However, requires centralized access to the full set, and thus impractical large-scale real-world applications, where too large be stored on one single machine. In paper, we propose distributed version of (DPOSS) with bounded guarantee....

Stochastic gradient descent (SGD) is a popular and efficient method with wide applications in training deep neural nets other nonconvex models. While the behavior of SGD well understood convex learning setting, existing theoretical results for applied to objective functions are far from mature. For example, require impose nontrivial assumption on uniform boundedness gradients all iterates encountered process, which hard verify practical implementations. In this paper, we establish rigorous...

In this paper, a video watermarking scheme is proposed to make the watermarked sequence robust against collusion attacks. Now temporal correlation between adjacent frames poses severe challenges for applications,. If same or redundant used embedding in every frame of video, can be estimated and then removed by watermark estimation attack., also if uncorrelated watermarks are frame, these washed out with filtering. The algorithm uses method high-precision motion compensation which presented...

Most heuristic algorithms for NP-hard combinatorial optimization problems require expertise in both the problem domains and methods. Recent research has begun to apply Deep Neural Network learning heuristics automatically. These works mainly focus with simple formulations, such as Travelling Salesman Problem Vehicle Routing defined on Euclidean graphs. This paper presents a novel deep reinforcement based algorithm Capacitated Arc which is more complex non-Euclidean information The proposed...

In this paper we propose a two-party certificate less authenticated key agreement scheme which has proven to be secure in the random oracle model. The new protocol is as long each party at least one uncompromised secret. addition, higher efficiency than several other strongly protocols(such Lippold[2], protocol[3], protocol[4]).

Purpose The purpose of this study is to develop a novel region-based convolutional neural networks (R-CNN) approach that more efficient while at least as accurate existing R-CNN methods. In way, the proposed method, namely R 2 -CNN, provides powerful tool for pedestrian extraction person re-identification, which involve huge number images and needs be extracted efficiently meet real-time requirement. Design/methodology/approach -CNN tested on two types data sets. first one USC Pedestrian...

DDS technology is an innovative circuit architecture. It has the advantages of low phase noise, fast frequency switching, small cubage, high resolutions. This paper presents a agile synthesizer with simple structure, anti-jamming, stray and noise.

Object detection from still images has been among the most active and challenging area in computer vision recently. In contrast, fully supervised object video rarely investigated. this paper, we propose an algorithm to improve performance of video. Our proposed method is based on empirical property that trajectory important for videos. We use filter outliers, determine probable location correct errors. compare our with baseline which regard frames as directly. Experiments show outperform...

R-CNN style methods are sorts of the state-of-the-art object detection methods, which consist region proposal generation and deep CNN classification. However, phase in this paradigm is usually time consuming, would slow down whole testing. This paper suggests that value discrepancies among features convolutional feature maps contain plenty useful spatial information, proposes a simple approach to extract information for fast The proposed method, namely Relief (R2-CNN), adopts novel generator...

Evolutionary Reinforcement Learning (ERL) that applying Algorithms (EAs) to optimize the weight parameters of Deep Neural Network (DNN) based policies has been widely regarded as an alternative traditional reinforcement learning methods. However, evaluation iteratively generated population usually requires a large amount computational time and can be prohibitively expensive, which may potentially restrict applicability ERL. Surrogate is often used reduce burden in EAs. Unfortunately, ERL,...