Right now, we round weights during discretization. This may not work well if there is a systematic bias in the weights, where e.g. many weights have a fractional component between 0 and 0.5, which would all get rounded down. Stochastic rounding should statistically eliminate this bias.
Right now, we round weights during discretization. This may not work well if there is a systematic bias in the weights, where e.g. many weights have a fractional component between 0 and 0.5, which would all get rounded down. Stochastic rounding should statistically eliminate this bias.