Large-Scale Learning with Less RAM via Randomization

Daniel Golovin, D. Sculley, Brendan McMahan, Michael Young
Proceedings of the 30th International Conference on Machine Learning, PMLR 28(2):325-333, 2013.

Abstract

We reduce the memory footprint of popular large-scale online learning methods by projecting our weight vector onto a coarse discrete set using randomized rounding. Compared to standard 32-bit float encodings, this reduces RAM usage by more than 50% during training and by up 95% when making predictions from a fixed model, with almost no loss in accuracy. We also show that randomized counting can be used to implement per-coordinate learning rates, improving model quality with little additional RAM. We prove these memory-saving methods achieve regret guarantees similar to their exact variants. Empirical evaluation confirms excellent performance, dominating standard approaches across memory versus accuracy tradeoffs.

Cite this Paper


BibTeX
@InProceedings{pmlr-v28-golovin13, title = {Large-Scale Learning with Less RAM via Randomization}, author = {Golovin, Daniel and Sculley, D. and McMahan, Brendan and Young, Michael}, booktitle = {Proceedings of the 30th International Conference on Machine Learning}, pages = {325--333}, year = {2013}, editor = {Dasgupta, Sanjoy and McAllester, David}, volume = {28}, number = {2}, series = {Proceedings of Machine Learning Research}, address = {Atlanta, Georgia, USA}, month = {17--19 Jun}, publisher = {PMLR}, pdf = {http://proceedings.mlr.press/v28/golovin13.pdf}, url = {https://proceedings.mlr.press/v28/golovin13.html}, abstract = {We reduce the memory footprint of popular large-scale online learning methods by projecting our weight vector onto a coarse discrete set using randomized rounding. Compared to standard 32-bit float encodings, this reduces RAM usage by more than 50% during training and by up 95% when making predictions from a fixed model, with almost no loss in accuracy. We also show that randomized counting can be used to implement per-coordinate learning rates, improving model quality with little additional RAM. We prove these memory-saving methods achieve regret guarantees similar to their exact variants. Empirical evaluation confirms excellent performance, dominating standard approaches across memory versus accuracy tradeoffs.} }
Endnote
%0 Conference Paper %T Large-Scale Learning with Less RAM via Randomization %A Daniel Golovin %A D. Sculley %A Brendan McMahan %A Michael Young %B Proceedings of the 30th International Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2013 %E Sanjoy Dasgupta %E David McAllester %F pmlr-v28-golovin13 %I PMLR %P 325--333 %U https://proceedings.mlr.press/v28/golovin13.html %V 28 %N 2 %X We reduce the memory footprint of popular large-scale online learning methods by projecting our weight vector onto a coarse discrete set using randomized rounding. Compared to standard 32-bit float encodings, this reduces RAM usage by more than 50% during training and by up 95% when making predictions from a fixed model, with almost no loss in accuracy. We also show that randomized counting can be used to implement per-coordinate learning rates, improving model quality with little additional RAM. We prove these memory-saving methods achieve regret guarantees similar to their exact variants. Empirical evaluation confirms excellent performance, dominating standard approaches across memory versus accuracy tradeoffs.
RIS
TY - CPAPER TI - Large-Scale Learning with Less RAM via Randomization AU - Daniel Golovin AU - D. Sculley AU - Brendan McMahan AU - Michael Young BT - Proceedings of the 30th International Conference on Machine Learning DA - 2013/05/13 ED - Sanjoy Dasgupta ED - David McAllester ID - pmlr-v28-golovin13 PB - PMLR DP - Proceedings of Machine Learning Research VL - 28 IS - 2 SP - 325 EP - 333 L1 - http://proceedings.mlr.press/v28/golovin13.pdf UR - https://proceedings.mlr.press/v28/golovin13.html AB - We reduce the memory footprint of popular large-scale online learning methods by projecting our weight vector onto a coarse discrete set using randomized rounding. Compared to standard 32-bit float encodings, this reduces RAM usage by more than 50% during training and by up 95% when making predictions from a fixed model, with almost no loss in accuracy. We also show that randomized counting can be used to implement per-coordinate learning rates, improving model quality with little additional RAM. We prove these memory-saving methods achieve regret guarantees similar to their exact variants. Empirical evaluation confirms excellent performance, dominating standard approaches across memory versus accuracy tradeoffs. ER -
APA
Golovin, D., Sculley, D., McMahan, B. & Young, M.. (2013). Large-Scale Learning with Less RAM via Randomization. Proceedings of the 30th International Conference on Machine Learning, in Proceedings of Machine Learning Research 28(2):325-333 Available from https://proceedings.mlr.press/v28/golovin13.html.

Related Material