Learning Symbolic Interactions for Interpretable State-Space Modeling
Primary Area: neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Keywords: Bayesian back-propagation, complex systems
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
Abstract: A general model to learn complex system dynamics will be helpful for us to understand how natural and computational networks of simple computation units solve complex problems. We formulate discrete event dynamics as a Bayesian neural network with skip connections: we use linearity to select hidden features to interact and combine the effects of these interactions, and we use nonlinearity (exponential and logarithm) to compound these interactions. To make learning scalable, we derive a Bayesian backpropagation algorithm that computes the expected loss gradient through propagating filtering and smoothing probabilities of hidden features. Experiments demonstrate that our algorithm can data-efficiently capture complex system dynamics in several fields with meaningful interactions.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 7500
Loading