Reviews: Training Deep Networks without Learning Rates Through Coin Betting

Summary: This paper is based on the notion (established in existing works) that subgradient descent can be interpreted as betting on a coin. It generalizes existing results to data-dependent bets and, consequently, data-dependent convergence bounds that improve upon the existing ones. The algorithm is then adapted for the training of neural networks and evaluated on this task experimentally. Quality: The derivations are mathematically sound. I verified the proof of Theorem 1. The changes made to adapt COCOB for the training of neural networks (Algo 1 -- Algo 2) are sensible and intuitive.