Convergence for Discrete Parameter Update Schemes

Paul Wilson, Fabio Zanasi, George Constantinides

Modern deep learning models require immense computational resources, motivating research into low-precision training. Quantised training addresses this by representing training com…