The implementation of Machine Learning (ML) algorithms on stand-alone small-scale devices allows the incorporation of new services and advanced functionalities without the need of resorting to remote com- puting systems. Despite having undeniable advantages with respect to conventional general-purpose devices, e.g. in terms of cost/performance ratios, small-scale systems suffer of issues related to their resource-limited nature, like limited battery capacity and processing power. In order to deal with such limitations, we propose to merge local Rademacher Complexities and bit-based hypothesis spaces to build thrifty models, which can be effectively implemented on small-scale resource-limited devices. Experiments, carried out on a smartphone in a Human Activity Recognition application, show the benefits of the proposed approach in terms of model accuracy and battery duration.
Learning with Few Bits on Small--Scale Devices: from Regularization to Energy Efficiency
Oneto Luca;
2014-01-01
Abstract
The implementation of Machine Learning (ML) algorithms on stand-alone small-scale devices allows the incorporation of new services and advanced functionalities without the need of resorting to remote com- puting systems. Despite having undeniable advantages with respect to conventional general-purpose devices, e.g. in terms of cost/performance ratios, small-scale systems suffer of issues related to their resource-limited nature, like limited battery capacity and processing power. In order to deal with such limitations, we propose to merge local Rademacher Complexities and bit-based hypothesis spaces to build thrifty models, which can be effectively implemented on small-scale resource-limited devices. Experiments, carried out on a smartphone in a Human Activity Recognition application, show the benefits of the proposed approach in terms of model accuracy and battery duration.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
