Cloud for High-Performance Statistical Learning
Big data asks for scalable algorithms, but scalability is just one aspect of the problem. Many applications also require the speedy processing of large volumes of data. Examples include supporting financial decision making, advanced services in digital libraries, mining medical data from magnetic resonance imaging, and also analyzing social media graphs. The velocity of machine learning is often boosted by deploying graphics processing units (GPUs) or distributed algorithms, but rarely both. We are developing high-performance supervised and unsupervised statistical learning algorithms that are accelerated on GPU clusters. The tools in our current scope are dimensionality reduction by random projection, low-dimensional embedding by self-organizing maps, clustering bynature-inspired methods, classification by support vector machines, and further methods are constantly being considered. While MapReduce is commonly used in distributed-memory machine learning, and there are MapReduce frameworks that can easily be made GPU-aware, and one that was specifically designed for distributed GPUs, we believe that higher performance can be achieved in a less restrictive communication model. This design decision should not affect the modularity of a data mining workflow, as given an efficient distributed storage model, the learning algorithms smoothly integrate with the rest of the processing steps, irrespective of whether they use the MapReduce model.
Since the cost of a GPU cluster is high and the deployment is far from being trivial, Cloud for High-Performance Statistical Learning (ChiP-SL) enables the verification, rapid dissemination, and quick adaptation of the algorithms being developed. Based on our past experience, such a solution is particularly relevant in organizations which do not normally require high computational power, such as digital libraries, or social and archaeological simulations. By benchmarking on cloud GPU instances, we may provide a turn-key solution for the instant deployment of high-performance machine learning algorithms by sharing the respective images of the virtual servers. This saves the trouble of getting the relevant libraries of the correct version number working, compiling the code, and setting up paths, making these solution more accessible for both academic and commercial users.
This work is supported by AWS in Education Machine Learning Grant award.