UC Santa Cruz

Distributed algorithms in machine learning follow two main flavors: horizontal partitioning, where the data is distributed across multiple slaves and vertical partitioning, where the model parameters are partitioned across multiple machines. The main drawback of the former strategy is that the model parameters need to be replicated on every machine. This is problematic when the number of parameters is very large, and hence cannot fit in a single machine. This drawback of the latter strategy is that the data needs to be replicated on each machine, thus failing to scale to massive datasets.

The goal of this thesis is to achieve the best of both worlds by partitioning both - the data as well as the model parameters, thus enabling the training of more sophisticated models on massive datasets. In order to do so, we exploit a structure that is observed in several machine learning models, which we term as \textit{Double-Separability}. Double-Separability basically means that the objective function of the model can be decomposed into independent sub-functions which can be computed independently. For distributed machine learning, this implies that both data and model parameters can partitioned across machines and stochastic updates for parameters can be carried out independently and without any locking. Furthermore, double-separability naturally lends itself to developing efficient asynchronous algorithms which enable computation and communication to happen in parallel, offering further speedup.

Some machine learning models such as Matrix Factorization directly exhibit double-separability in their objective function, however the majority of models do not. My work explores techniques to reformulate the objective function of such models to cast them into double-separable form. Often this involves introducing additional auxiliary variables that have nice interpretations. In this direction, I have developed Hybrid Parallel algorithms for machine learning tasks that include {\it Latent Collaborative Retrieval}, {\it Multinomial Logistic Regression}, {\it Variational Inference for Mixture of Exponential Families} and {\it Factorization Machines}. The software resulting from this work are available for public use under an open-source license.