arXiv subjects

Natasha Sharygina

Publications and source records attributed to Natasha Sharygina.

6 recordsLinked to original sources

arXiv2026

Neuron Activation-based Computation of Logical Explanations for Deep Neural Networks

Formal explainability of classifying neural networks (NNs) is an active area of research, providing explanations with provable guarantees of the classification within continuous regions of the input feature space. However, the existing techniques are either limited to individual input features without guarantees on their relations or the provided solutions fail to scale to deep architectures. This paper addresses these issues by introducing a flexible symbolic framework for an efficient, guided computation of explanations of the NN behavior, parametrized by the activations of internal neurons, and using logical engines such as SMT solvers. Unlike prior methods that rely on specialized NN verifiers, our method yields explanations that are not restricted in shape. Our algorithm is implementable on top of a general-purpose logical solver, isolating the NN-specific encoding from the algorithmic framework. We experimented with a wide range of benchmarks from the domains of image recognition and medicine, illustrating the advantages of the new method, particularly in computational efficiency. Notably, our approach enables logical explanation of deep networks not amenable to prior logic-based methods.

cs.LO↗

arXiv2026

Termination analysis with interpolation-based transition invariant generation

Termination and nontermination of infinite-state systems are complementary problems that, despite their close connection, are typically addressed by separate techniques. The core idea of this paper is to connect termination and nontermination analysis, enabling the two to share intermediate results and guide one another. We present a new termination analysis approach based on the generation of well-founded transition invariants. It leverages Craig interpolation for transition invariant generation, capturing the structural reasons for termination. The proposed technique extends safety-based nontermination analysis, enabling it to prove both termination and nontermination within a unified framework. We implemented our approach in the GOLEM verification framework and evaluated it on benchmarks from the Termination Competition (TermComp). Empirical results demonstrate that combining termination and nontermination is beneficial and yields performance comparable to state-of-the-art tools.

cs.LO↗

arXiv2025

Space Explanations of Neural Network Classification

We present a novel logic-based concept called Space Explanations for classifying neural networks that gives provable guarantees of the behavior of the network in continuous areas of the input feature space. To automatically generate space explanations, we leverage a range of flexible Craig interpolation algorithms and unsatisfiable core generation. Based on real-life case studies, ranging from small to medium to large size, we demonstrate that the generated explanations are more meaningful than those computed by state-of-the-art.

cs.LG↗

arXiv2014

Monotonic Abstraction Techniques: from Parametric to Software Model Checking

Monotonic abstraction is a technique introduced in model checking parameterized distributed systems in order to cope with transitions containing global conditions within guards. The technique has been re-interpreted in a declarative setting in previous papers of ours and applied to the verification of fault tolerant systems under the so-called "stopping failures" model. The declarative reinterpretation consists in logical techniques (quantifier relativizations and, especially, quantifier instantiations) making sense in a broader context. In fact, we recently showed that such techniques can over-approximate array accelerations, so that they can be employed as a meaningful (and practically effective) component of CEGAR loops in software model checking too.

cs.LO↗

arXiv2013

Abstraction and Acceleration in SMT-based Model-Checking for Array Programs

Abstraction (in its various forms) is a powerful established technique in model-checking; still, when unbounded data-structures are concerned, it cannot always cope with divergence phenomena in a satisfactory way. Acceleration is an approach which is widely used to avoid divergence, but it has been applied mostly to integer programs. This paper addresses the problem of accelerating transition relations for unbounded arrays with the ultimate goal of avoiding divergence during reachability analysis of abstract programs. For this, we first design a format to compute accelerations in this domain; then we show how to adapt the so-called 'monotonic abstraction' technique to efficiently handle complex formulas with nested quantifiers generated by the acceleration preprocessing. Notably, our technique can be easily plugged-in into abstraction/refinement loops, and strongly contributes to avoid divergence: experiments conducted with the MCMT model checker attest the effectiveness of our approach on programs with unbounded arrays, where acceleration and abstraction/refinement technologies fail if applied alone.

cs.LO↗

arXiv2013

Interpolation Properties and SAT-based Model Checking

Craig interpolation is a widespread method in verification, with important applications such as Predicate Abstraction, CounterExample Guided Abstraction Refinement and Lazy Abstraction With Interpolants. Most state-of-the-art model checking techniques based on interpolation require collections of interpolants to satisfy particular properties, to which we refer as "collectives"; they do not hold in general for all interpolation systems and have to be established for each particular system and verification environment. Nevertheless, no systematic approach exists that correlates the individual interpolation systems and compares the necessary collectives. This paper proposes a uniform framework, which encompasses (and generalizes) the most common collectives exploited in verification. We use it for a systematic study of the collectives and of the constraints they pose on propositional interpolation systems used in SAT-based model checking.

cs.LO↗