Efficient Use of Limited-Memory Resources to Accelerate Linear Learning

We propose a generic algorithmic building block to accelerate training of machine learning models on heterogenous compute systems. The scheme allows to efficiently employ compute accelerators such as GPUs and FPGAs for the training of large-scale machine learning models, when the training data exceeds their memory capacity. Also, it provides adaptivity to any system's memory hierarchy in terms of size and processing speed. Our technique builds upon primal-dual coordinate methods, and uses duality gap information to dynamically decide which part of the data should be made available for fast processing. We provide a strong theoretical motivation for our gap-based selection scheme and provide an efficient practical implementation thereof. To illustrate the power of our approach we demonstrate its performance for training of generalized linear models on large scale datasets exceeding the memory size of a modern GPU, showing an order-of-magnitude speedup over existing approaches.