Big Learning with Little RAM

In large-scale machine learning, available memory (RAM) is often a key constraint, both during model training and when making new predictions. In this paper, we reduce memory cost by projecting our weight vector β ∈ R onto a coarse discrete set using randomized rounding. Because the values of the discrete set can be stored more compactly than standard 32-bit float encodings, this reduces RAM usage by 50% during training and by up 90% at prediction time. Theoretical analysis provides safety guarantees that bound the regret added by this projection. Empirical evaluation confirms excellent results in practice, adding only an additional 0.01% to logistic loss in testing.