Asynchronous Stochastic Gradient Research Articles

Bagua

Recent years have witnessed a growing list of systems for distributed data-parallel training. Existing systems largely fit into two paradigms, i.e., parameter server and MPI-style collective operations. On the algorithmic side, researchers have proposed a wide range of techniques to lower the communication via "system relaxations": quantization, decentralization, and communication delay. However, most, if not all, existing systems only rely on standard synchronous and asynchronous stochastic gradient (SG) based optimization, therefore, cannot take advantage of all possible optimizations that the machine learning community has been developing recently. Given this emerging gap between the current landscapes of systems and theory, we build Bagua, a MPI-style communication library, providing a collection of primitives, that is both flexible and modular to support state-of-the-art system relaxation techniques of distributed training. Powered by this design, Bagua has a great ability to implement and extend various state-of-the-art distributed learning algorithms. In a production cluster with up to 16 machines (128 GPUs), Bagua can outperform PyTorch-DDP, Horovod and BytePS in the end-to-end training time by a significant margin (up to 2X) across a diverse range of tasks. Moreover, we conduct a rigorous tradeoff exploration showing that different algorithms and system relaxations achieve the best performance over different network conditions.

Distributed stochastic gradient descent for link prediction in signed social networks

This paper considers the link prediction problem defined over a signed social network, where the relationship between any two network users can be either positive (friends) or negative (foes). Given a portion of the relationships, the goal of link prediction is to identify the rest unknown ones. This task resorts to completing the adjacency matrix of the signed social network, which is low rank or approximately low rank. Considering the large scale of the adjacency matrix, in this paper, we adopt low-rank matrix factorization models for the link prediction problem and solve them through asynchronous distributed stochastic gradient descent algorithms. The low-rank matrix factorization models effectively reduce the size of the parameter space, while the asynchronous distributed stochastic gradient descent algorithms enable fast completion of the adjacency matrix. We validate the proposed algorithms using two real-world datasets on a distributed shared-memory computation platform. Numerical results demonstrate that the asynchronous distributed stochastic gradient descent algorithms achieve nearly linear computional speedups with respect to the number of computational threads, and are able to complete an adjacency matrix of ten billions of entries within 10 s.

Asynchronous self-organizing maps

A recently defined energy function which leads to a self-organizing map is used as a foundation for an asynchronous neural-network algorithm. We generalize the existing stochastic gradient approach to an asynchronous parallel stochastic gradient method for generating a topological map on a distributed computer system (MIMD). A convergence proof is presented and simulation results on a set of problems are included. A practical problem using the energy function approach is that a summation over the entire network is required during the computation of updates. Using simulations we demonstrate effective algorithms that use efficient sampling for the approximation of these sums.