DiNNO: Distributed Neural Network Optimization for Multi-Robot Collaborative Learning
DiNNO: Distributed Neural Network Optimization for Multi-Robot Collaborative Learning
复制标题
DOI：
10.1109/lra.2022.3142402
复制
发表时间：
2021-09
影响因子：
5.2
通讯作者：
Javier Yu;Joseph A. Vincent;M. Schwager
中科院分区：
文献类型：
--
作者：
Javier Yu;Joseph A. Vincent;M. Schwager
我们提出了DiNNO，一种分布式算法，使一组机器人能够在网状网络上通信时协同优化深度神经网络模型。每个机器人只能访问自己的数据，并维护自己版本的神经网络，但最终学会的模型就像在所有数据上集中训练一样好。没有机器人通过无线网络发送原始数据，保护数据隐私并确保有效利用无线带宽。在每次迭代中，每个机器人都近似地优化一个增广拉格朗日函数，然后将得到的权重传递给相邻机器人，更新对偶变量，并进行重复。最终，所有机器人的局部模型权值达到一致。对于凸目标函数，这种一致性是全局最优的。与许多现有的方法不同，我们在机器人相关的深度学习任务上测试我们的算法，这些任务具有重要的模型架构。我们将DiNNO与两个基准分布式深度学习算法在(i) MNIST图像分类任务，（ii）多机器人隐式映射任务和（iii）多机器人强化学习任务中进行了比较。在这些实验中，我们证明了DiNNO在面对非凸深度学习目标、时变通信图和流数据时表现良好。在所有的实验中，我们的方法都优于基线，并且能够达到与集中训练模型相当的验证损失。请参阅msl.stanford.edu/projects/dist_nn_train获取视频和代码。
We present DiNNO, a distributed algorithm that enables a group of robots to collaboratively optimize a deep neural network model while communicating over a mesh network. Each robot only has access to its own data and maintains its own version of the neural network, but eventually learns a model that is as good as if it had been trained on all the data centrally. No robot sends raw data over the wireless network, preserving data privacy and ensuring efficient use of wireless bandwidth. At each iteration, each robot approximately optimizes an augmented Lagrangian function, then communicates the resulting weights to its neighbors, updates dual variables, and repeats. Eventually, all robots' local model weights reach a consensus. For convex objective functions, this consensus is a global optimum. Unlike many existing methods we test our algorithm on robotics related, deep learning tasks with nontrivial model architectures. We compare DiNNO to two benchmark distributed deep learning algorithms in (i) an MNIST image classification task, (ii) a multi-robot implicit mapping task, and (iii) a multi-robot reinforcement learning task. In these experiments we show that DiNNO performs well when faced with nonconvex deep learning objectives, time varying communication graphs, and streaming data. In all experiments our method outperforms baselines, and was able to achieve validation loss equivalent to centrally trained models. See msl.stanford.edu/projects/dist_nn_train for videos and code.