Probing Deep Learning Models by Logic
Public support
Provider
Czech Science Foundation
Programme
Standard projects
Call for proposals
SGA0202600001
Main participants
České vysoké učení technické v Praze / Fakulta elektrotechnická
Contest type
VS - Public tender
Contract ID
26-23758S
Alternative language
Project name in Czech
Probing Deep Learning Models by Logic
Annotation in Czech
Deep learning may be paving the way toward general artificial intelligence. While its progress is impressive, it is also concerning that we are expected to place blind trust in models that function as black boxes—networks of billions of weights, where their creators could potentially embed hidden and undesirable “surprises”. In this project, we will contribute to the overall goal of making deep learning models more transparent by developing tools for probing and benchmarking them. For that we will use first-order logic and non-monotonic logics. We propose to contribute to the state of the art along the following axes: (i) Evaluating deep learning models and/or the algorithms used to train them, and (ii) Extracting the implicit logical knowledge in these models and reasoning with that knowledge. Our focus will be on two major model types: graph neural networks (GNNs) and large language models (LLMs).
Scientific branches
R&D category
ZV - Basic research
OECD FORD - main branch
10201 - Computer sciences, information science, bioinformathics (hardware development to be 2.2, social aspect to be 5.8)
OECD FORD - secondary branch
—
OECD FORD - another secondary branch
—
CEP - equivalent branches <br>(according to the <a href="http://www.vyzkum.cz/storage/att/E6EF7938F0E854BAE520AC119FB22E8D/Prevodnik_oboru_Frascati.pdf">converter</a>)
AF - Documentation, librarianship, work with information<br>BC - Theory and management systems<br>BD - Information theory<br>IN - Informatics
Solution timeline
Realization period - beginning
Jan 1, 2026
Realization period - end
Dec 31, 2028
Project status
Z - Beginning multi-year project
Latest support payment
—
Data delivery to CEP
Confidentiality
S - Úplné a pravdivé údaje o projektu nepodléhají ochraně podle zvláštních právních předpisů
Data delivery code
CEP26-GA0-GA-R
Data delivery date
May 6, 2026
Finance
Total approved costs
10,356 thou. CZK
Public financial support
10,356 thou. CZK
Other public sources
0 thou. CZK
Non public and foreign sources
0 thou. CZK