9 problems

Effectiveness of Local SGD. Local SGD can dominate mini-batch SGD over this problem class in the regime when $\tau, QB \mathbb{L} H$ and $\boldsymbol{\sigma},\zeta_\star$ are small…

The paper studies inductive proofs for the effect of delays on iterative reinforcement-learning algorithms and compares them with robustness results for temporal-difference learnin…

Let $\mathcal{G}$ be a network whose strong-robustness property is used to ensure resilience in distributed hypothesis testing and estimation. Strong-robustness conjecture. This gr…

Consider problems in which the globally optimal parameter belongs to a finite set, and suppose the problems admit notions of statistical heterogeneity and distributed sequential de…

Decentralized stochastic gradient descent (D-SGD) is an optimization method in which agents communicate over a network and use stochastic gradients; data heterogeneity refers to di…

Let $M>2$ and let $N$ be the number of channels. For a problem instance $(M,N,\boldsymbol\mu)$, consider all possible permutations of arms in the set $\mathcal{S}(M)$ and the round…

Sparse estimation tradeoff conjecture. If some protocol estimates the mean for any distribution $P\in\mathcal{P}_s$ with mean-squared loss $R$ and communication cost $C$, then

Communication–loss tradeoff conjecture. The tradeoff between communication and squared loss demonstrated by this protocol is essentially optimal up to a logarithmic factor.

Complete-network conjecture. A complete network achieves a lower steady state MSD in the learning process than other networks.