DISCD: Distributed Lossy Semantic Communication for Logical Deduction of Hypothesis
Ahmet Faruk Saz, Siheng Xiong, Faramarz Fekri
TL;DR
This work tackles distributed hypothesis deduction where each node holds only partial State of the World (SotW) and communication is constrained. It introduces DISCD, a lossy semantic communication framework that uses inductive logical probabilities and cont-information to select semantically informative messages exchanged with a central server, which updates a global SotW and broadcasts informative updates. The approach provides a formal convergence analysis, including PAC-style guarantees, and shows empirically that semantically informed transmissions yield faster and more reliable hypothesis deduction with reduced communication overhead. The contribution enables bandwidth-efficient, privacy-preserving coordinated reasoning across distributed agents and has practical implications for real-time, explainable deduction in networks with partial observability. Overall, DISCD demonstrates that strategically chosen semantic messages can guide distributed reasoning toward the true state more efficiently than random transmissions.
Abstract
In this paper, we address hypothesis testing in a distributed network of nodes, where each node has only partial information about the State of the World (SotW) and is tasked with determining which hypothesis, among a given set, is most supported by the data available within the node. However, due to each node's limited perspective of the SotW, individual nodes cannot reliably determine the most supported hypothesis independently. To overcome this limitation, nodes must exchange information via an intermediate server. Our objective is to introduce a novel distributed lossy semantic communication framework designed to minimize each node's uncertainty about the SotW while operating under limited communication budget. In each communication round, nodes determine the most content-informative message to send to the server. The server aggregates incoming messages from all nodes, updates its view of the SotW, and transmits back the most semantically informative message. We demonstrate that transmitting semantically most informative messages enables convergence toward the true distribution over the state space, improving deductive reasoning performance under communication constraints. For experimental evaluation, we construct a dataset designed for logical deduction of hypotheses and compare our approach against random message selection. Results validate the effectiveness of our semantic communication framework, showing significant improvements in nodes' understanding of the SotW for hypothesis testing, with reduced communication overhead.