Local Message Passing on Frustrated Systems - UAI Oral Session 6

Explore a 25-minute conference talk from the Uncertainty in Artificial Intelligence event that delves into an innovative approach to message passing on factor graphs with cyclic structures. Learn about the limitations of the sum-product algorithm (SPA) on graphs with small cycles and discover a novel alternative that challenges the extrinsic principle of SPA. Understand how replacing the local SPA message update rule with an optimized generic mapping leads to improved performance while maintaining simplicity. Examine the evaluation of this method on two classes of cyclic graphs: the 2x2 fully connected Ising grid and factor graphs for symbol detection on linear communication channels with inter-symbol interference. Gain insights into a new loss function inspired by the Bethe approximation, enabling unsupervised training for large-scale practical applications. Access the presentation slides to visualize key concepts and findings from this cutting-edge research in probabilistic inference and message passing algorithms.