Distributed Inference over Unreliable Channels

As modern AI workloads outgrow traditional centralized computing architectures, distributing the computing tasks is becoming attractive for performance and scaling. However, this promise comes with reliability challenges due to unreliable hardware and noisy communication channels affecting the participating nodes. Thus, in this work we pursue methods for protecting distributed-inference tasks from channel noise in communication and storage. For two central inference tasks: binary classification and regression, our contributions enhance the reliability of the state-of-the-art ensemble-learning methods called bagging and boosting. This is achieved by a variety of results: modeling, analytical, algorithmic, and empirical, whose core is a post-training algorithm for weighting the individual predictors in the ensemble based on both their training loss and the parameters of their channels.

Ph.D. student Under the supervision of Prof. Yuval Cassuto.