Unified approach combining gradient descent and multiplicative updates.
problem Combining gradient descent and multiplicative updates for machine learning.
method Introduces hypentropy and a family of matrix-based updates.
result Derives tight regret bounds for the new family of updates.