On Hardware-Aware Probabilistic Frameworks for Resource Constrained Embedded Applications

Edge reasoning attempts to mitigate latency and privacy shortcomings of cloud computing paradigms. However, it introduces additional challenges linked to the devices' resource constraints and the applications' dynamic conditions. To address these challenges, we have proposed a hardware-aware probabilistic framework that optimizes the target machine learning model under actual hardware constraints. This framework relies on tractable probabilistic models, as they facilitate efficient inference, while exhibiting a number of traits relevant to the application range of interest: robustness to missing data, joint prediction capabilities, explainability, and small data needs. In this work, we expand on this framework by introducing a discriminative-generative approach to model learning, which retains the robustness of a generative model under missing data but can potentially improve its discriminative performance. In addition, we demonstrate how the applicability of this framework goes beyond classification tasks, and can be used for density estimation tasks, relevant to applications such as mobile speaker verification.