Modern communication networks have become very complicated and highly dynamic, which makes them hard to model, predict control. In this paper, we develop a novel experience-driven approach that can learn well control network from its own experience rather than an accurate mathematical just as human learns new skill (such driving, swimming, etc). Specifically, we, for the first time, propose leverage emerging Deep Reinforcement Learning (DRL) enabling model-free in networks; present effective...

In this paper, we propose to leverage the emerging deep learning techniques for spatiotemporal modeling and prediction in cellular networks, based on big system data. First, perform a preliminary analysis dataset from China Mobile, use traffic load as an example show non-zero temporal autocorrelation spatial correlation among neighboring Base Stations (BSs), which motivate us discover both dependencies our study. Then present hybrid model prediction, includes novel autoencoder-based Long...

Cloud Radio Access Networks (RANs) have become a key enabling technique for the next generation (5G) wireless communications, which can meet requirements of massively growing data traffic. However, resource allocation in cloud RANs still needs to be further improved order reach objective minimizing power consumption and meeting demands users over long operational period. Inspired by success Deep Reinforcement Learning (DRL) on solving complicated control problems, we present novel DRL-based...

Automatic decision-making approaches, such as reinforcement learning (RL), have been applied to (partially) solve the resource allocation problem adaptively in cloud computing system. However, a complete framework exhibits high dimensions state and action spaces, which prohibit usefulness of traditional RL techniques. In addition, power consumption has become one critical concerns design control systems, degrades system reliability increases cooling cost. An effective dynamic management...

Structured weight pruning is a representative model compression technique of DNNs to reduce the storage and computation requirements accelerate inference. An automatic hyperparameter determination process necessary due large number flexible hyperparameters. This work proposes AutoCompress, an structured framework with following key performance improvements: (i) effectively incorporate combination schemes in process; (ii) adopt state-of-art ADMM-based as core algorithm, propose innovative...

In this paper, we aim to study networking problems from a whole new perspective by leveraging emerging deep learning, develop an experience-driven approach, which enables network or protocol learn the best way control itself its own experience (e.g., runtime statistics data), just as human learns skill. We present design, implementation and evaluation of reinforcement learning (DRL)-based framework, DRL-CC (DRL for Congestion Control), realizes our design philosophy on multi-path TCP (MPTCP)...

Mobile crowdsourcing (MCS) is now an important source of information for smart cities, especially with the help unmanned aerial vehicles (UAVs) and driverless cars. They are equipped different kinds high-precision sensors, can be scheduled/controlled completely during data collection, which will make MCS system more robust. However, they limited to energy constraint, long-term, long-distance sensing tasks, cities almost too crowded set stationary charging station. Towards this end, in paper...

In this paper, we propose the first deep reinforcement learning framework to estimate optimal Dynamic Treatment Regimes from observational medical data. This is more flexible and adaptive for high dimensional action state spaces than existing methods model real life complexity in heterogeneous disease progression treatment choices, with goal provide doctor patients data-driven personalized decision recommendations. The proposed contains a supervised step predict most possible expert actions;...

The raw depth image captured by the indoor sen-sor usually has an extensive range of missing values due to inherent limitations such as inability perceive transparent objects and limited distance range. incomplete map burdens many downstream vision tasks, a rising number completion methods have been proposed alleviate this issue. While most existing meth-ods can generate accurate dense maps from sparse uniformly sampled maps, they are not suitable for complementing large contiguous regions...

Despite the prominent success of general object detection, performance and efficiency Small Object Detection (SOD) are still unsatisfactory. Unlike existing works that struggle to balance tradeoff between inference speed SOD performance, in this paper, we propose a novel Scale-aware Knowledge Distillation (ScaleKD), which transfers knowledge complex teacher model compact student model. We design two modules boost quality transfer distillation for SOD: 1) scale-decoupled feature module...

Automatic decision-making approaches, such as reinforcement learning (RL), have been applied to (partially) solve the resource allocation problem adaptively in cloud computing system. However, a complete framework exhibits high dimensions state and action spaces, which prohibit usefulness of traditional RL techniques. In addition, power consumption has become one critical concerns design control systems, degrades system reliability increases cooling cost. An effective dynamic management...

Fine-grained instance segmentation is considerably more complicated and challenging than semantic segmentation. Most existing methods only focus on accuracy without paying much attention to inference latency, which, critical real-time applications, such as autonomous driving. In this paper, we aim bridge the gap between by presenting a novel model for segmentation, Sem2Ins, which effectively generates boundaries according leveraging conditional generative adversarial networks (cGANs) coupled...

In this paper, we focus on general-purpose Distributed Stream Data Processing Systems (DSDPSs) , which deal with processing of unbounded streams continuous data at scale distributedly in real or near-real time. A fundamental problem a DSDPS is the scheduling (i.e., assigning workload to workers/machines) objective minimizing average end-to-end tuple widely-used solution distribute evenly over machines cluster round-robin manner, obviously not efficient due lack consideration for...

Vision-based autonomous urban driving in dense traffic is quite challenging due to the complicated environment and dynamics of behaviors. Widely-applied methods either heavily rely on hand-crafted rules or learn from limited human experience, which makes them hard generalize rare but critical scenarios. In this paper, we present a novel CAscade Deep REinforcement learning framework, CADRE, achieve model-free vision-based driving. derive representative latent features raw observations, first...

In this paper, we focus on general-purpose Distributed Stream Data Processing Systems (DSDPSs), which deal with processing of unbounded streams continuous data at scale distributedly in real or near-real time. A fundamental problem a DSDPS is the scheduling (i.e., assigning workload to workers/machines) objective minimizing average end-to-end tuple widely-used solution distribute evenly over machines cluster round-robin manner, obviously not efficient due lack consideration for communication...

In this paper, we focus on general-purpose Distributed Stream Data Processing Systems (DSDPSs), which deal with processing of unbounded streams continuous data at scale distributedly in real or near-real time. A fundamental problem a DSDPS is the scheduling objective minimizing average end-to-end tuple widely-used solution to distribute workload evenly over machines cluster round-robin manner, obviously not efficient due lack consideration for communication delay. Model-based approaches do...

Meta-learning enables a model to learn from very limited data undertake new task. In this paper, we study the general meta-learning with adversarial samples. We present algorithm, ADML (ADversarial Meta-Learner), which leverages clean and samples optimize initialization of learning in an manner. leads following desirable properties: 1) it turns out be effective even cases only samples; 2) is robust samples, i.e., unlike other algorithms, minor performance degradation when there are 3) sheds...

This paper presents the deep reinforcement learning (DRL) framework to estimate optimal Dynamic Treatment Regimes from observational medical data. is more flexible and adaptive for high dimensional action state spaces than existing methods model real-life complexity in heterogeneous disease progression treatment choices, with goal of providing doctors patients data-driven personalized decision recommendations. The proposed DRL comprises (i) a supervised step predict expert actions, (ii)...

Channel pruning can effectively reduce both computational cost and memory footprint of the original network while keeping a comparable accuracy performance. Though great success has been achieved in channel for 2D image-based convolutional networks (CNNs), existing works seldom extend methods to 3D point-based neural (PNNs). Directly implementing CNN PNNs undermine performance because different representations images point clouds as well architecture disparity. In this paper, we proposed CP...

Modern communication networks have become very complicated and highly dynamic, which makes them hard to model, predict control. In this paper, we develop a novel experience-driven approach that can learn well control network from its own experience rather than an accurate mathematical just as human learns new skill (such driving, swimming, etc). Specifically, we, for the first time, propose leverage emerging Deep Reinforcement Learning (DRL) enabling model-free in networks; present effective...

Video question answering (VideoQA) is a very important but challenging multimedia task, which automatically analyzes questions and videos generates accurate answers. However, research on VideoQA still in its infancy. In this article, we propose novel memory augmented deep recurrent neural network (MA-DRNN) model for VideoQA, features new method encoding questions, augmentation using the emerging differentiable computer (DNC). Specifically, encode textual (questions) information before visual...

Experience-driven networking has emerged as a new and highly effective approach for resource allocation in complex communication networks. Deep Reinforcement Learning (DRL) been shown to be useful technique enabling experience-driven networking. In this paper, we focus on practical fundamental problem networking: when network configurations are changed, how train DRL agent effectively quickly adapt the environment. We present an Actor-Critic-based Transfer learning framework Traffic...