To seed discussion, I will attempt to organize research efforts in large-scale machine learning by looking at common computational problems across all of machine learning, and the challenges of creating efficient parallel algorithms for them. I'll begin by identifying four common types of computational bottlenecks that occur across all of machine learning, or prototype algorithmic problems: N-body problems, graph operations, linear algebra, and optimization. Within each category, I'll discuss what we can or cannot learn from the existing body of work in scientific computing, highlight a few of the most successful and recent specific serial algorithms that have been developed for concreteness, and discuss what makes them easy or hard to parallelize. I'll synthesize some of these observations to obtain a list of desiderata for parallel machine learning algorithms research and software toolkits.