Towards Hardware-Aware Tractable Learning of Probabilistic Models

Smart portable applications increasingly rely on edge computing due to privacyand latency concerns. But guaranteeing always-on functionality comes with twomajor challenges: heavily resource-constrained hardware; and dynamic applicationconditions. Probabilistic models present an ideal solution to these challenges:they are robust to missing data, allow for joint predictions and have small dataneeds. In addition, ongoing efforts in the field of tractable learning have resultedin probabilistic models with strict inference efficiency guarantees. However, thecurrent notions of tractability are often limited to model complexity, disregardingthe hardware’s specifications and constraints. We propose a novel resource-awarecost metric that takes into consideration the hardware’s properties in determiningwhether the inference task can be efficiently deployed. We use this metric toevaluate the performance versus resource trade-off relevant to the applicationof interest, and we propose a strategy that selects the device settings that canoptimally meet users’ requirements. We showcase our framework on a mobileactivity recognition scenario, and on a variety of benchmark datasets representativeof the field of tractable learning and of the applications of interest.status: accepte