As the big data paradigm is gaining momentum, learning algorithms trained through fast stochastic gradient descent methods are becoming the de-facto standard in the industry world. Still, even these simple procedures cannot be used completely "off-the-shelf" because parameters, e.g. the learning rate, has to be properly tuned to the particular problem to achieve fast convergence.
The online learning framework is a powerful tool to design fast learning algorithms able to work in both the stochastic and adversarial setting.
In this talk I will introduce new advancements in the time-varying regularization framework for online learning, that allows to derive almost parameter-free adaptive algorithms. In particular, I will focus on a new algorithm based on a dimension-free exponentiated gradient. Contrary to the existing online algorithms, it achieves an optimal regret bound, up to logarithmic terms, without any parameter nor any prior knowledge about the optimal solution.