CEO Neural Network： A Consistency-Aware Fusion Method for Semantic Multi-Observation Communication

As the Chinese proverb says， “Three cobblers together can rival Zhuge Liang，” meaning that collective wisdom often surpasses individual genius. This concept is equally applicable in the research of semantic communication. To address the challenges posed by multi-source observations involving multiple categories of coexisting information as well as diverse and easily disturbed semantic expressions in real-world scenarios， this paper focuses on the Semantic Multi-Observation Communication System （SMOCS） and proposes a neural network architecture based on a semantic consistency fusion mechanism， aiming to realize effective integration and reliable discrimination of semantic information from multiple observational perspectives. The system simulates different observers modeling the same semantic event through multiple structurally heterogeneous pre-trained neural networks， each independently generating semantic discrimination probabilities， thereby forming a rich and diverse set of semantic expressions. Accordingly， this paper introduces a Chief Executive Officer （CEO） fusion network as the semantic integration module， which focuses on mining latent distribution consistency and complementary structures from multi-source discrimination results. This network does not directly participate in semantic feature extraction； rather， through a deep fusion mechanism， it learns the intrinsic correlations and semantic aggregation methods from the outputs of multiple observation models， thus producing more stable and accurate final decisions. The entire system’s structure reflects a semantic communication optimization strategy driven by “multi-observation， collaborative modeling， and unified decision-making.” To verify the efficacy of the proposed method， the MNIST handwritten digit recognition task was selected as the experimental platform， where Multilayer Perceptron （MLP）， Convolutional Neural Network （CNN）， Recurrent Neural Network （RNN）， Long Short-Term Memory （LSTM） Network， and Residual Neural Network （ResNet） subnetworks were pre-trained with fixed weights， and only the CEO network was trained to ensure the purity of the information fusion phase. Experimental results demonstrate that， compared to single-neural-network classifiers， the proposed system achieves significant improvements in recognition accuracy， robustness， and anti-interference capability， exhibiting stronger fault tolerance and expression stability when faced with semantic perturbations and information loss. To verify the generalizability of the proposed architecture， simulation experiments were conducted on a text semantic transmission task， demonstrating the applicability and superior performance of the model across multimodal and diverse scenarios.