We gratefully acknowledge support from
the Simons Foundation and member institutions.
Full-text links:


Current browse context:


Change to browse by:

References & Citations

DBLP - CS Bibliography


(what is this?)
CiteULike logo BibSonomy logo Mendeley logo del.icio.us logo Digg logo Reddit logo ScienceWISE logo

Computer Science > Machine Learning

Title: Does Symbolic Knowledge Prevent Adversarial Fooling?

Authors: Stefano Teso
Abstract: Arguments in favor of injecting symbolic knowledge into neural architectures abound. When done right, constraining a sub-symbolic model can substantially improve its performance and sample complexity and prevent it from predicting invalid configurations. Focusing on deep probabilistic (logical) graphical models -- i.e., constrained joint distributions whose parameters are determined (in part) by neural nets based on low-level inputs -- we draw attention to an elementary but unintended consequence of symbolic knowledge: that the resulting constraints can propagate the negative effects of adversarial examples.
Comments: Short position paper. Accepted at the Ninth International Workshop on Statistical Relational AI (StarIA 2020)
Subjects: Machine Learning (cs.LG); Machine Learning (stat.ML)
Cite as: arXiv:1912.10834 [cs.LG]
  (or arXiv:1912.10834v1 [cs.LG] for this version)

Submission history

From: Stefano Teso [view email]
[v1] Thu, 19 Dec 2019 11:50:33 GMT (17kb)

Link back to: arXiv, form interface, contact.