Reviews: Federated Multi-Task Learning

The paper generalizes an existing framework for distributed multitask learning (COCOA) and extends it to handle practical systems challenges such as communications cost, stragglers, etc. The proposed technique (MOCHA) makes the framework robust and faster. Pros: - Rigorous theoretical analysis - Sound experiment methodology Cons: - Not very novel (compared to [47]) and just a mild variation over [47] including the analysis. The Equation 5 depends on \Delta \alpha_t* (minimizer of sub-problem). My understanding is that this minimizer is not known and cannot be computed accurately at every node, and therefore, we need the approximation.