ParMAC: distributed optimisation of nested functions, with application to learning binary autoencoders

Many powerful machine learning models are based on the composition of multiple processing layers, such as deep nets, which gives rise to nonconvex objective functions. A general, recent approach to optimise such “nested” functions is the method of auxiliary coordinates (MAC) (Carreira-Perpiñán and Wang, 2014). MAC introduces an auxiliary coordinate for each data point in order to decouple the nested model into independent submodels. This decomposes the optimisation into steps that alternate between training single layers and updating the coordinates. It has the advantage that it reuses existing singlelayer algorithms, introduces parallelism, and does not need to use chain-rule gradients, so it works with nondifferentiable layers. With large-scale problems, or when distributing the computation is necessary for faster training, the dataset may not fit in a single machine. It is then essential to limit the amount of communication between machines so it does not obliterate the benefit of parallelism. We describe a general way to achieve this, ParMAC. ParMAC works on a cluster of processing machines with a circular topology and alternates two steps until convergence: one step trains the submodels in parallel using stochastic updates, and the other trains the coordinates in parallel. Only submodel parameters, no data or coordinates, are ever communicated between machines. ParMAC exhibits high parallelism, low communication overhead, and facilitates data shuffling, load balancing, fault tolerance and streaming data processing. We study the convergence of ParMAC and propose a theoretical model of its runtime and parallel speedup. To illustrate our general results in a specific algorithm, we develop ParMAC to learn binary autoencoders with application to fast, approximate image retrieval. We implement this using Message Passing Interface (MPI) in a distributed system and demonstrate nearly perfect speedups in a 128-processor cluster with a training set of 100 million high-dimensional points. The speedups achieved agree well with the prediction of our theoretical speedup model.