Search

We investigate challenges and possibilities of formal reasoning for encoder-only transformers (EOT), meaning sound and complete methods for verifying or interpreting behaviour. In detail, we condense related formal reasoning tasks in the form of a naturally occurring satisfiability problem (SAT). We find that SAT is undecidable if we consider EOT, commonly considered in the expressiveness community. Furthermore, we identify practical scenarios where SAT is decidable and establish corresponding complexity bounds. Besides trivial cases, we find that quantized EOT, namely those restricted by some fixed-width arithmetic, lead to the decidability of SAT due to their limited attention capabilities. However, the problem remains difficult, as we establish those scenarios where SAT is NEXPTIME-hard and those where we can show that it is solvable in NEXPTIME for quantized EOT. To complement our theoretical results, we put our findings and their implications in the overall perspective of formal reasoning.

Verifying properties and interpreting the behaviour of deep neural networks (DNN) is an important task given their ubiquitous use in applications, including safety-critical ones, and their black-box nature. We propose an automata-theoric approach to tackling problems arising in DNN analysis. We show that the input-output behaviour of a DNN can be captured precisely by a (special) weak B\"uchi automaton and we show how these can be used to address common verification and interpretation tasks of DNN like adversarial robustness or minimum sufficient reasons.

Otto's Theorem characterises the bisimulation-invariant PTIME queries over graphs as exactly those that can be formulated in the polyadic mu-calculus, hinging on the Immerman-Vardi Theorem which characterises PTIME (over ordered structures) by First-Order Logic with least fixpoints. This connection has been extended to characterise bisimulation-invariant EXPTIME by an extension of the polyadic mu-calculus with functions on predicates, making use of Immerman's characterisation of EXPTIME by Second-Order Logic with least fixpoints. In this paper we show that the bisimulation-invariant versions of all classes in the exponential time hierarchy have logical counterparts which arise as extensions of the polyadic mu-calculus by higher-order functions. This makes use of the characterisation of k-EXPTIME by Higher-Order Logic (of order k+1) with least fixpoints, due to Freire and Martins., Comment: In Proceedings GandALF 2022, arXiv:2209.09333

Output reachability and adversarial robustness are among the most relevant safety properties of neural networks. We show that in the context of Message Passing Neural Networks (MPNN), a common Graph Neural Network (GNN) model, formal verification is impossible. In particular, we show that output reachability of graph-classifier MPNN, working over graphs of unbounded size, non-trivial degree and sufficiently expressive node labels, cannot be verified formally: there is no algorithm that answers correctly (with yes or no), given an MPNN, whether there exists some valid input to the MPNN such that the corresponding output satisfies a given specification. However, we also show that output reachability and adversarial robustness of node-classifier MPNN can be verified formally when a limit on the degree of input graphs is given a priori. We discuss the implications of these results, for the purpose of obtaining a complete picture of the principle possibility to formally verify GNN, depending on the expressiveness of the involved GNN models and input-output specifications.

We investigate the complexity of the reachability problem for (deep) neural networks: does it compute valid output given some valid input? It was recently claimed that the problem is NP-complete for general neural networks and specifications over the input/output dimension given by conjunctions of linear inequalities. We recapitulate the proof and repair some flaws in the original upper and lower bound proofs. Motivated by the general result, we show that NP-hardness already holds for restricted classes of simple specifications and neural networks. Allowing for a single hidden layer and an output dimension of one as well as neural networks with just one negative, zero and one positive weight or bias is sufficient to ensure NP-hardness. Additionally, we give a thorough discussion and outlook of possible extensions for this direction of research on neural network verification., Comment: arXiv admin note: substantial text overlap with arXiv:2108.13179

We investigate the complexity of the reachability problem for (deep) neural networks: does it compute valid output given some valid input? It was recently claimed that the problem is NP-complete for general neural networks and conjunctive input/output specifications. We repair some flaws in the original upper and lower bound proofs. We then show that NP-hardness already holds for restricted classes of simple specifications and neural networks with just one layer, as well as neural networks with minimal requirements on the occurring parameters.

We investigate the expressive power of the two main kinds of program logics for complex, non-regular program properties found in the literature: those extending propositional dynamic logic (PDL), and those extending the modal mu-calculus. This is inspired by the recent discovery of a decidable program logic called Visibly Pushdown Fixpoint Logic with Chop which extends both the modal mu-calculus and PDL over visibly pushdown languages, which, so far, constituted the ends of two pillars of decidable program logics. Here we show that this logic is not only more expressive than either of its two fragments, but in fact even more expressive than their union. Hence, the decidability border amongst program logics has been properly pushed up. We complete the picture by providing results separating all the PDL-based and modal fixpoint logics with regular, visibly pushdown and arbitrary context-free constructions., Comment: In Proceedings EXPRESS/SOS 2021, arXiv:2108.09624

Local fixpoint iteration describes a technique that restricts fixpoint iteration in function spaces to needed arguments only. It has been studied well for first-order functions in abstract interpretation and also in model checking. Here we consider the problem for least and greatest fixpoints of arbitrary type order. We define an abstract algebra of simply typed higher-order functions with fixpoints that can express fixpoint evaluation problems as they occur routinely in various applications, including program verification. We present an algorithm that realises local fixpoint iteration for such higher-order fixpoints, prove its correctness and study its optimisation potential in the context of several applications., Comment: In Proceedings GandALF 2020, arXiv:2009.09360

The sequent calculus is a formalism for proving validity of statements formulated in First-Order Logic. It is routinely used in computer science modules on mathematical logic. Formal proofs in the sequent calculus are finite trees obtained by successively applying proof rules to formulas, thus simplifying them step-by-step. Students often struggle with the mathematical formalities and the level of abstraction that topics like formal logic and formal proofs involve. The difficulties can be categorised as syntactic or semantic. On the syntactic level, students need to understand what a correctly formed proof is, how rules can be applied (on paper for instance) without leaving the mathematical framework of the sequent calculus, and so on. Beyond this, on the semantic level, students need to acquire strategies that let them find the right proof. The Sequent Calculus Trainer is a tool that is designed to aid students in learning the techniques of proving given statements formally. In this paper we describe the didactical motivation behind the tool and the techniques used to address issues on the syntactic as well as on the semantic level., Comment: In Proceedings ThEdu'17, arXiv:1803.00722

We study the following natural variation on the classical universality problem: given a language $L(M)$ represented by $M$ (e.g., a DFA/RE/NFA/PDA), does there exist an integer $\ell \geq 0$ such that $\Sigma^\ell \subseteq L(M)$? In the case of an NFA, we show that this problem is NEXPTIME-complete, and the smallest such $\ell$ can be doubly exponential in the number of states. This particular case was formulated as an open problem in 2009, and our solution uses a novel and involved construction. In the case of a PDA, we show that it is recursively unsolvable, while the smallest such $\ell$ is not bounded by any computable function of the number of states. In the case of a DFA, we show that the problem is NP-complete, and $e^{\sqrt{n \log n} (1+o(1))}$ is an asymptotically tight upper bound for the smallest such $\ell$, where $n$ is the number of states. Finally, we prove that in all these cases, the problem becomes computationally easier when the length $\ell$ is also given in binary in the input: it is polynomially solvable for a DFA, PSPACE-complete for an NFA, and co-NEXPTIME-complete for a PDA., Comment: This is the full version of the conference paper https://doi.org/10.4230/LIPIcs.STACS.2020.16

We consider simulation games played between Spoiler and Duplicator on two B\"uchi automata in which the choices made by Spoiler can be buffered by Duplicator in several buffers before she executes them on her structure. We show that the simulation games are useful to approximate the inclusion of trace closures of languages accepted by finite-state automata, which is known to be undecidable. We study the decidability and complexity and show that the game with bounded buffers can be decided in polynomial time, whereas the game with one unbounded and one bounded buffer is highly undecidable. We also show some sufficient conditions on the automata for Duplicator to win the game (with unbounded buffers)., Comment: In Proceedings GandALF 2016, arXiv:1609.03648

We study the axiomatisability of the iteration-free fragment of Propositional Dynamic Logic with Intersection and Tests. The combination of program composition, intersection and tests makes its proof-theory rather difficult. We develop a normal form for formulae which minimises the interaction between these operators, as well as a refined canonical model construction. From these we derive an axiom system and a proof of its strong completeness., Comment: In Proceedings GandALF 2016, arXiv:1609.03648