NeSy-MM

Our most powerful AI agents rely almost exclusively on neural networks, which cannot easily be told to obey certain rules or incorporate background knowledge. Relational Neurosymbolic Markov Models (NeSy-MMs) combine the learning abilities of neural networks with the guarantees of symbolic reasoning, applied to sequential decision-making under uncertainty.

NeSy-MMs integrate neurosymbolic AI with Markov models by using probabilistic models over relations as their core representation. Unlike standard hidden Markov models that use unstructured latent variables, NeSy-MMs decompose symbolic relations over time, similar to how planning algorithms work, while learning unknown dynamics with neural networks.

This gives NeSy-MMs three key abilities that other models lack:

1. Consistent generation. NeSy-MMs generate sequences that respect given constraints (e.g., navigation rules), while deep Markov models and visual transformers trained on the same data fail to do so.
2. Out-of-distribution generalisation. Trained on small grids with one enemy, NeSy-MMs generalise to larger grids, more enemies, and longer time horizons, unlike transformer baselines.
3. Test-time interventions. New constraints can be added at test time without retraining (e.g., forbidding the agent from entering part of the room).

Deep Markov Model

NeSy-MM (Ours)

Visual Transformer

resources

• Paper: arXiv
• Code: ML-KULeuven/nesy-mm
• Blog post: AIHub