Abstract
The interpretation of logical expressions into loss functions has given rise to so-called differentiable logics. They function as a bridge between formal logic and machine learning, offering a novel approach for property-driven training. The added expressiveness of these logics comes at the price of a more intricate semantics for first-order quantifiers. To ease their integration into machine-learning backends, we explore how to formalize semantics for first-order differentiable logics using the Mathematical Components library in the Rocq proof assistant. We seek to give rigorous semantics for quantifiers, verify their properties with respect to other logical connectives, as well as prove the soundness and completeness of the resulting logics.
| Original language | English |
|---|---|
| Title of host publication | 8th International Symposium on AI Verification (SAIV 2025), Zagreb, Croatia, July 21--22, 2025 |
| Number of pages | 12 |
| Publication date | 2025 |
| Pages | 1-12 |
| Publication status | Published - 2025 |
| Event | International Symposium on AI Verification - Zagreb, Croatia Duration: 21 Jul 2025 → 22 Jul 2025 Conference number: 8 https://aiverification.org/2025/ |
Symposium
| Symposium | International Symposium on AI Verification |
|---|---|
| Number | 8 |
| Country/Territory | Croatia |
| City | Zagreb |
| Period | 21/07/2025 → 22/07/2025 |
| Internet address |
Keywords
- Neural Network Verification
- Formal Specifications
- Loss Functions
- Differentiable Logics
- Interactive Theorem Proving
Fingerprint
Dive into the research topics of 'Quantifiers for Differentiable Logics in Rocq (Extended Abstract)'. Together they form a unique fingerprint.Cite this
- APA
- Author
- BIBTEX
- Harvard
- Standard
- RIS
- Vancouver