The Minimax Complexity of Distributed Optimization
This work addresses the theoretical limits of distributed optimization algorithms for researchers and practitioners in machine learning, though it is incremental in extending existing frameworks.
The thesis tackles the problem of distributed stochastic optimization in the intermittent communication setting, establishing matching upper and lower bounds for convex and non-convex objectives, with simple SGD variants proven optimal.
In this thesis, I study the minimax oracle complexity of distributed stochastic optimization. First, I present the "graph oracle model", an extension of the classic oracle complexity framework that can be applied to study distributed optimization algorithms. Next, I describe a general approach to proving optimization lower bounds for arbitrary randomized algorithms (as opposed to more restricted classes of algorithms, e.g., deterministic or "zero-respecting" algorithms), which is used extensively throughout the thesis. For the remainder of the thesis, I focus on the specific case of the "intermittent communication setting", where multiple computing devices work in parallel with limited communication amongst themselves. In this setting, I analyze the theoretical properties of the popular Local Stochastic Gradient Descent (SGD) algorithm in convex setting, both for homogeneous and heterogeneous objectives. I provide the first guarantees for Local SGD that improve over simple baseline methods, but show that Local SGD is not optimal in general. In pursuit of optimal methods in the intermittent communication setting, I then show matching upper and lower bounds for the intermittent communication setting with homogeneous convex, heterogeneous convex, and homogeneous non-convex objectives. These upper bounds are attained by simple variants of SGD which are therefore optimal. Finally, I discuss several additional assumptions about the objective or more powerful oracles that might be exploitable in order to develop better intermittent communication algorithms with better guarantees than our lower bounds allow.