DiNNO: Distributed Neural Network Optimization for Multi-Robot Collaborative Learning

We present DiNNO, a distributed algorithm that enables group of robots to collaboratively optimize deep neural network model while communicating over mesh network. Each robot only has access its own data and maintains version the network, but eventually learns is as good if it had been trained on all centrally. No sends raw wireless preserving privacy ensuring efficient use bandwidth. At each iteration, approximately optimizes an augmented Lagrangian function, then communicates resulting weights neighbors, updates dual variables, repeats. Eventually, robots’ local reach consensus. For convex objective functions, this consensus global optimum. Unlike many existing methods we test our robotics-related, learning tasks with nontrivial architectures. compare DiNNO two benchmark algorithms in (i) MNIST image classification task, (ii) multi-robot implicit mapping (iii) reinforcement task. In these experiments show performs well when faced nonconvex objectives, time-varying communication graphs, streaming data. method outperforms baselines, was able achieve validation loss equivalent centrally models. See msl.stanford.edu/projects/dist_nn_train for videos code.