More Like this

1. Communication-efficient ADMM using quantization-aware Gaussian process regression

   https://doi.org/10.1016/j.ejco.2024.100098  

   Duarte, Aldo; Nghiem, Truong X; Wei, Shuangqing (September 2024, EURO Journal on Computational Optimization)

   In networks consisting of agents communicating with a central coordinator and working together to solve a global optimization problem in a distributed manner, the agents are often required to solve private proximal minimization subproblems. Such a setting often requires a decomposition method to solve the global distributed problem, resulting in extensive communication overhead. In networks where communication is expensive, it is crucial to reduce the communication overhead of the distributed optimization scheme. Gaussian processes (GPs) are effective at learning the agents' local proximal operators, thereby reducing the communication between the agents and the coordinator. We propose combining this learning method with adaptive uniform quantization for a hybrid approach that can achieve further communication reduction. In our approach, due to data quantization, the GP algorithm is modified to account for the introduced quantization noise statistics. We further improve our approach by introducing an orthogonalization process to the quantizer's input to address the inherent correlation of the input components. We also use dithering to ensure uncorrelation between the quantizer's introduced noise and its input. We propose multiple measures to quantify the trade-off between the communication cost reduction and the optimization solution's accuracy/optimality. Under such metrics, our proposed algorithms can achieve significant communication reduction for distributed optimization with acceptable accuracy, even at low quantization resolutions. This result is demonstrated by simulations of a distributed sharing problem with quadratic cost functions for the agents. 

   more » « less
2. Communication-efficient federated hypergradient computation via aggregated iterative differentiation

   Xiao, P; Ji, K (July 2023, International Conference on Machine Learning)

   Federated bilevel optimization has attracted increasing attention due to emerging machine learning and communication applications. The biggest challenge lies in computing the gradient of the upper-level objective function (ie, hypergradient) in the federated setting due to the nonlinear and distributed construction of a series of global Hessian matrices. In this paper, we propose a novel communication-efficient federated hypergradient estimator via aggregated iterative differentiation (AggITD). AggITD is simple to implement and significantly reduces the communication cost by conducting the federated hypergradient estimation and the lower-level optimization simultaneously. We show that the proposed AggITD-based algorithm achieves the same sample complexity as existing approximate implicit differentiation (AID)-based approaches with much fewer communication rounds in the presence of data heterogeneity. Our results also shed light on the great advantage of ITD over AID in the federated/distributed hypergradient estimation. This differs from the comparison in the non-distributed bilevel optimization, where ITD is less efficient than AID. Our extensive experiments demonstrate the great effectiveness and communication efficiency of the proposed method. 

   more » « less
3. ILMPQ : An Intra-Layer Multi-Precision Deep Neural Network Quantization framework for FPGA

   Chang, Sung-En; Li, Yanyu; Sun, Mengshu; Wang, Yanzhi; Lin, Xue (February 2021, The Fifth Workshop on Cognitive Architectures (CogArch 2021))

   This work targets the commonly used FPGA (field-programmable gate array) devices as the hardware platform for DNN edge computing. We focus on DNN quantization as the main model compression technique. The novelty of this work is: We use a quantization method that supports multiple precisions along the intra-layer dimension, while the existing quantization methods apply multi-precision quantization along the inter-layer dimension. The intra-layer multi-precision method can uniform the hardware configurations for different layers to reduce computation overhead and at the same time preserve the model accuracy as the inter-layer approach. Our proposed ILMPQ DNN quantization framework achieves 70.73% Top1 accuracy in ResNet-18 on the ImageNet dataset. We also validate the proposed MSP framework on two FPGA devices i.e., Xilinx XC7Z020 and XC7Z045. We achieve 3.65× speedup in end-to-end inference time on the ImageNet, comparing with the fixed-point quantization method. 

   more » « less
4. Learning Proximal Operators with Gaussian Process and Adaptive Quantization in Distributed Optimization

   https://doi.org/10.31390/gradschool_dissertations.6484  

   Duarte, Aldo (May 2024, LSU Doctoral Dissertations)

   In networks consisting of agents communicating with a central coordinator and working together to solve a global optimization problem in a distributed manner, the agents are often required to solve private proximal minimization subproblems. Such a setting often requires a further decomposition method to solve the global distributed problem, resulting in extensive communication overhead. In networks where communication is expensive, it is crucial to reduce the communication overhead of the distributed optimization scheme. Integrating Gaussian processes (GP) as a learning component to the Alternating Direction Method of Multipliers (ADMM) has proven effective in learning each agent's local proximal operator to reduce the required communication exchange. In this work, we propose to combine this learning method with adaptive uniform quantization in a hybrid approach that can achieve further communication reduction when solving a distributed optimization problem with ADMM. This adaptive quantization first considers setting the mid-value and window length according to the mean and covariance given by GP. In a later stage of our study, this adaptation is extended to also consider the variation of the quantization bit resolution. In addition, a convergence analysis of this setting is derived, leading to convergence conditions and error bounds in the cases where convergence cannot be formally proven. Furthermore, we study the impact of the communication decision-making of the coordinator, leading to the proposition of several query strategies using the agent's uncertainty measures given by the regression process. Extensive numerical experiments of a distributed sharing problem with quadratic cost functions for the agents have been conducted throughout this study. The results have demonstrated that the various algorithms proposed have successfully achieved their primary goal of minimizing the overall communication overhead while ensuring that the global solutions maintain satisfactory levels of accuracy. The favorable accuracy observed in the numerical experiments is consistent with the findings of the derived convergence analysis. In instances where convergence proof is lacking, we have shown that the overall ADMM residual remains bounded by a diminishing threshold. This implies that we can anticipate our algorithmic solutions to closely approximate the actual solution, thus validating the reliability of our approaches. 

   more » « less
5. On the Convergence of Communication-Efficient Local SGD for Federated Learning

   Hongchang Gao, An Xu (February 2021, 35th AAAI Conference on Artificial Intelligence (AAAI 2021))

   Federated Learning (FL) has attracted increasing attention in recent years. A leading training algorithm in FL is local SGD, which updates the model parameter on each worker and averages model parameters across different workers only once in a while. Although it has fewer communication rounds than the classical parallel SGD, local SGD still has large communication overhead in each communication round for large machine learning models, such as deep neural networks. To address this issue, we propose a new communicationefficient distributed SGD method, which can significantly reduce the communication cost by the error-compensated double compression mechanism. Under the non-convex setting, our theoretical results show that our approach has better communication complexity than existing methods and enjoys the same linear speedup regarding the number of workers as the full-precision local SGD. Moreover, we propose a communication-efficient distributed SGD with momentum, which also has better communication complexity than existing methods and enjoys a linear speedup with respect to the number of workers. At last, extensive experiments are conducted to verify the performance of our proposed methods. 

   more » « less