Despite the growing influence of AI in society, experts in industrial settings face challenges in the integration of humans into cyber physical systems, in order to contribute human-like reasoning to machine learning processes. While classical graph neural networks (GNNs) operate on predefined graph structures and more recent approaches attempt to infer connectivity from data, both paradigms typically require large labeled datasets to achieve robust generalization. This paper addresses these challenges by proposing the Laplacian Associative-Projective Neural Network (LAPNN), a brain-inspired vector-symbolic architecture for learning spatio-temporal structural motifs in human cyber physical networks. We evaluate the ability of the LAPNN to generalize across human subjects in human-machine collaborative tasks, using minimal training data from a modified version of the collaborative action (CoAx) dataset. We use a cross-subject, incremental ablation with a theoretical early stopping threshold, to determine the number of human subject datasets necessary for accurate graph-level classification. The LAPNN demonstrates strong data-efficient generalization at graph-level classification when compared to a temporal hypergraph neural network baseline. In addition to graph-level tasks, the LAPNN natively supports node-level analogical-style retrieval and mapping, enabling the identification of part–whole relationships and is-a class membership within learned motifs. This capability is a feature of the model architecture itself, rather than emerging from auxiliary decoders. The results suggest that the LAPNN enables data-efficient graph learning for human cyber-physical networks, with strong generalization under limited supervision, providing a viable alternative to embedding-based GNNs for cognitive human cyber-physical system design.
Gish et al. (Sat,) studied this question.