Mobile Edge Artificial Intelligence: Opportunities and Challenges

Description

Mobile Edge Artificial Intelligence: Opportunities and Challenges presents recent advances in wireless technologies and nonconvex optimization techniques for designing efficient edge AI systems. The book includes comprehensive coverage on modeling, algorithm design and theoretical analysis. Through typical examples, the powerfulness of this set of systems and algorithms is demonstrated, along with their abilities to make low-latency, reliable and private intelligent decisions at network edge. With the availability of massive datasets, high performance computing platforms, sophisticated algorithms and software toolkits, AI has achieved remarkable success in many application domains. As such, intelligent wireless networks will be designed to leverage advanced wireless communications and mobile computing technologies to support AI-enabled applications at various edge mobile devices with limited communication, computation, hardware and energy resources. Front Cover Mobile Edge Artificial Intelligence Copyright Contents List of figures Biography Yuanming Shi Kai Yang Zhanpeng Yang Yong Zhou Preface Acknowledgments Part 1 Introduction and overview 1 Motivations and organization 1.1 Motivations 1.2 Organization References 2 Primer on artificial intelligence 2.1 Basics of machine learning 2.1.1 Supervised learning 2.1.1.1 Logistic regression 2.1.1.2 Support vector machine 2.1.1.3 Decision tree 2.1.1.4 k-Nearest neighbors method 2.1.1.5 Neural network 2.1.2 Unsupervised learning 2.1.2.1 k-Means algorithm 2.1.2.2 Principal component analysis 2.1.2.3 Autoencoder 2.1.3 Reinforcement learning 2.1.3.1 Q-learning 2.1.3.2 Policy gradient 2.2 Models of deep learning 2.2.1 Convolutional neural network 2.2.2 Recurrent neural network 2.2.3 Graph neural network 2.2.4 Generative adversarial network 2.3 Summary References 3 Convex optimization 3.1 First-order methods 3.1.1 Gradient method for unconstrained problems 3.1.2 Gradient method for constrained problems 3.1.3 Subgradient descent method 3.1.4 Mirror descent method 3.1.5 Proximal gradient method 3.1.6 Accelerated gradient method 3.1.7 Smoothing for nonsmooth optimization 3.1.8 Dual and primal-dual methods 3.1.9 Alternating direction method of multipliers 3.1.10 Stochastic gradient method 3.2 Second-order methods 3.2.1 Newton's method 3.2.2 Quasi-Newton method 3.2.3 Gauss–Newton method 3.2.4 Natural gradient method 3.3 Summary References 4 Mobile edge AI 4.1 Overview 4.2 Edge inference 4.2.1 On-device inference 4.2.2 Edge inference via computation offloading 4.2.2.1 Server-based edge inference 4.2.2.2 Device-edge joint inference 4.3 Edge training 4.3.1 Data partition-based edge training 4.3.1.1 Distributed mode 4.3.1.2 Decentralized mode 4.3.2 Model partition-based edge training 4.4 Coded computing 4.5 Summary References Part 2 Edge inference 5 Model compression for on-device inference 5.1 Background on model compression 5.2 Layerwise network pruning 5.2.1 Problem statement 5.2.2 Convex approach for sparse objective and constraints 5.3 Nonconvex network pruning method with log-sum approximation 5.3.1 Log-sum approximation for sparse optimization 5.3.2 Iteratively reweighed minimization for log-sum approximation 5.4 Simulation results 5.4.1 Handwritten digits classification 5.4.2 Image classification 5.4.3 Keyword spotting inference 5.5 Summary References 6 Coded computing for on-device cooperative inference 6.1 Background on MapReduce 6.2 A communication-efficient data shuffling scheme 6.2.1 Communication model 6.2.2 Achievable data rates and DoF 6.3 A low-rank optimization framework for communication-efficient data shuffling 6.3.1 Interference alignment conditions 6.3.2 Low-rank optimization approach 6.4 Numerical algorithms 6.4.1 Nuclear norm relaxation 6.4.2 Iteratively reweighted least squares 6.4.3 Difference-of-convex (DC) programming approach 6.4.4 Computationally efficient DC approach 6.5 Simulation results 6.5.1 Convergence behaviors 6.5.2 Achievable DoF over local storage size 6.5.3 Scalability 6.6 Summary References 7 Computation offloading for edge cooperative inference 7.1 Background 7.1.1 Computation offloading 7.1.2 Edge inference via computation offloading 7.2 Energy-efficient wireless cooperative transmission for edge inference 7.2.1 Communication model 7.2.2 Power consumption model 7.2.3 Channel uncertainty model 7.2.4 Problem formulation 7.3 Computationally tractable approximation for probabilistic QoS constraints 7.3.1 Analysis of probabilistic QoS constraints 7.3.2 Scenario generation approach 7.3.3 Stochastic programming approach 7.3.4 Statistical learning-based robust optimization approach 7.3.4.1 Robust optimization approximation for probabilistic QoS constraints 7.3.4.2 Statistical learning approach for the high-probability region 7.3.4.2.1 Shape learning 7.3.4.2.2 Size calibration 7.3.4.3 Problem reformulation for problem P7.1.RO 7.3.5 A cost-effective channel sampling strategy 7.4 Reweighted power minimization approach with DC regularization 7.4.1 Nonconvex quadratic constraints 7.4.2 Reweighted power minimization with DC regularization 7.5 Simulation results 7.5.1 Benefits of considering CSI uncertainty 7.5.2 Advantages of overcoming the overconservativeness 7.5.3 Total power consumption 7.6 Summary References Part 3 Edge training 8 Over-the-air computation for federated learning 8.1 Background of federated learning and over-the-air computation 8.1.1 Federated learning 8.1.2 Over-the-air computation 8.2 System model 8.3 Fast model aggregation via over-the-air computation 8.3.1 A simple single-antenna case 8.3.2 Over-the-air computation for model aggregation with a multiantenna BS 8.3.3 Problem formulation 8.4 Sparse and low-rank optimization framework 8.5 Numerical algorithms 8.5.1 Convex relaxation approach 8.5.2 Iteratively reweighted minimization approach 8.5.3 DC programming approach 8.5.3.1 DC representations 8.5.3.2 DC program framework 8.6 Simulation results 8.6.1 Number of selected devices under MSE requirement 8.6.2 Performance of training an SVM classifier 8.7 Summary References 9 Reconfigurable intelligent surface aided federated learning 9.1 Background on reconfigurable intelligent surface 9.2 RIS empowered on-device distributed federated learning 9.2.1 System model 9.2.2 Problem formulation 9.3 Sparse and low-rank optimization framework 9.3.1 Two-step framework for sparse objective function 9.3.2 Alternating low-rank optimization for nonconvex biquadratic constraints 9.3.3 DC program for rank-one constraints 9.4 Simulation results 9.4.1 Device selection 9.4.2 Performance of federated learning 9.5 Summary References 10 Blind over-the-air computation for federated learning 10.1 Blind over-the-air computation 10.2 Problem formulation 10.3 Wirtinger flow algorithm for blind over-the-air computation 10.3.1 Wirtinger flow 10.3.2 Initialization strategies 10.4 Numerical results 10.5 Summary References Part 4 Final part: conclusions and future directions 11 Conclusions and future directions 11.1 Conclusions 11.2 Discussions and future directions Index Back Cover