Improved Bounds on Minimax Regret under Logarithmic Loss via Self-Concordance

no code implementations • ICML 2020 • Blair Bilodeau, Dylan Foster, Daniel Roy

We study the classical problem of forecasting under logarithmic loss while competing against an arbitrary class of experts.

Generalization via Derandomization

no code implementations • ICML 2020 • Jeffrey Negrea, Daniel Roy, Gintare Karolina Dziugaite

At the same time, we bound the risk of h^ in terms of a surrogate that is constructed by conditioning and shown to belong to a nonrandom class with uniformly small generalization error.

Adaptive Gradient Quantization for Data-Parallel SGD

1 code implementation • NeurIPS 2020 • Fartash Faghri, Iman Tabrizian, Ilia Markov, Dan Alistarh, Daniel Roy, Ali Ramezani-Kebrya

Many communication-efficient variants of SGD use gradient quantization schemes.

Quantization