… for parameter-free SCO are based on online parameter-free regret … is a novel parameter-free
certificate for SGD step size choice, … result that assumes no a-priori bounds on SGD iterates. …

… SGD results in a fully parameter-free method with convergence rate that matches optimally
tuned SGD up … to design a simple and efficient parameter-free algorithm, provided access to …

… There are two possible ways to use a SGD algorithm: to optimize a batch objective function,
… in turn makes SVMs the method of choice”. Indeed, in this paper we give a parameter-free …

… We believe that this difference is crucial in making DOG parameter-free, while the above-mentioned
method have a learning-rate parameter to tune (though Bernstein et al. (2020) report …

… show tuning-free matching of SGD is possible and achieved by several … Let us make our
setting more specific. We consider the … An algorithm A is called parameter-free if it achieves the …

… better than SGD, AdaGrad, and … make infinitely many infinitesimal SGD updates for each
sampled loss. We provide two improvements: making the approach work for a parameter-free up…

… w is destined to be pruned for some reason, instead of having a new parameter to denote
"to-prune", PDP lets SGD gradually make w itself smaller relatively against other parameters in …

… requirement makes the method not parameter-free. … The simplest way to make an algorithm
parameter-free is the … DoG is SGD’s best friend: A parameter-free dynamic step size schedule…

… on parameterfree adaptive gradient methods1 remains relatively limited. Specifically, most of
the existing parameter-free algorithms are essentially variants of SGD… This factor makes the …

… It is natural to ask whether one can achieve parameter-free … paper, and in Table 1 we make
the comparison between our … We conduct a simple grid search for D-SGD, D-SGD and D-…