We learn explicit temporal, causal, and logical structures from data so that model decisions can be interpreted, checked, revised, and reused as formal knowledge.
This direction connects neural representation learning with signal temporal logic, event and causal graphs, formal languages, and human knowledge. Explanation is not appended after prediction; symbolic structure is embedded in the model and optimisation problem.
The resulting representations support classification, requirement mining, semantic interpretation, verification, diagnosis, and controller synthesis across cyber-physical and industrial systems.
Temporal formulas, causal graphs, event structures, and program-like representations combine discrete topology with thresholds, time intervals, and weights. We study differentiable, sparse, structured, and search-based methods that optimise both without erasing semantics.
A symbol is useful only when it corresponds to observable evidence and supports consistent reasoning. We study mechanism-anchored predicates, monotone and causal structures, formal-language constraints, and human-in-the-loop refinement.
Noise, incomplete labels, uncertain dynamics, and domain changes can invalidate brittle rules. We develop robust semantics, uncertainty-aware inference, abstraction refinement, and verification interfaces that make the validity boundary explicit.
Time series, events, demonstrations, physical constraints, and expert concepts
Sparse attention, differentiable logic, causal graph learning, and system abstraction
Temporal formulas, event graphs, specifications, rules, and contracts
Explanation, requirement mining, semantic classification, and reusable guarantees
