Differentiable Logics

The Theory Behind the Practice

Differentiable logics are the engine that powers this entire project. The idea: take a logical formula expressing a desired property (e.g., “the classifier is robust to small perturbations”) and interpret it as a differentiable function over real numbers. This function can then serve as a loss term during neural network training — the optimizer simultaneously fits the data and satisfies the logical specification.

But the theory is still young. Current differentiable logics have limitations in expressivity, numerical stability, and the guarantees they provide. Extending them is both a mathematical and an engineering challenge.

What You’ll Work On

This theme is for students who enjoy the interplay between theory and implementation. You’ll work on the foundations of differentiable logics and their realization in tools like Vehicle.

Possible thesis directions:

What You’ll Learn

Project Results

property-driven-ml

A general open-source framework for property-driven machine learning: logical constraints over hyper-rectangle input regions are compiled into training loss terms through differentiable logics (Gödel, Łukasiewicz, Reichenbach, Yager, DL2, and Signal Temporal Logic), combined with PGD and Auto-PGD adversarial training. Gusts Grinbergs is a contributor to the framework, which is described in A General Framework for Property-Driven Machine Learning (Flinkow et al., 2025).

Signal Temporal Logic for Vehicle

An extension of the Vehicle specification language with Signal Temporal Logic operators (globally, finally, until) and closed-loop controller specifications, implemented as a built-in differentiable logic on top of STLCG++. Joint work by Alessandro Bruni, Matthew L. Daggitt, Gusts Gustavs Grīnbergs, and Ekaterina Komendantskaya; a paper on this work is to appear at LOPSTR+PPDP 2026.

Relevant Literature

Supervisors: Giorgio Bacci (AAU), Alessandro Bruni (ITU) – see Team