Neuro-symbolic programs, i.e. programs containing both machine learning components and traditional symbolic code, are becoming increasingly widespread. Finding a general methodology for verifying such programs is challenging due to both the number of different tools involved and the intricate interface between the "neural" and "symbolic" program components. In this paper we present a general decomposition of the neuro-symbolic verification problem into parts, and examine the problem of the embedding gap that occurs when one tries to combine proofs about the neural and symbolic components. To address this problem we then introduce Vehicle - standing as an abbreviation for a "verification condition language" - an intermediate programming language interface between machine learning frameworks, automated theorem provers, and dependently-typed formalisations of neuro-symbolic programs. Vehicle allows users to specify the properties of the neural components of neuro-symbolic programs once, and then safely compile the specification to each interface using a tailored typing and compilation procedure. We give a high-level overview of Vehicle’s overall design, its interfaces and compilation & type-checking procedures, and then demonstrate its utility by formally verifying the safety of a simple autonomous car controlled by a neural network, operating in a stochastic environment with imperfect information.

Abstract This paper serves as a comprehensive system description of version 2.0 of the Marabou framework for formal analysis of neural networks. We discuss the tool’s architectural design and highlight the major features and components introduced since its initial release.

Modern verification tools frequently rely on compiling high-level specifications to SMT queries. However, the high-level specification language is usually more expressive than the available solvers and therefore some syntactically valid specifications must be rejected by the tool. In such cases, the challenge is to provide a comprehensible error message to the user that relates the original syntactic form of the specification to the semantic reason it has been rejected.

This paper introduces Hypersequent GV (HGV), a modular and extensible core calculus for functional programming with session types that enjoys deadlock freedom, confluence, and strong normalisation. HGV exploits hyper-environments, which are collections of type environments, to ensure that structural congruence is type preserving. As a consequence we obtain an operational correspondence between HGV and HCP -- a process calculus based on hypersequents and in a propositions-as-types correspondence with classical linear logic (CLL). Our translations from HGV to HCP and vice-versa both preserve and reflect reduction. HGV scales smoothly to support Girard's Mix rule, a crucial ingredient for channel forwarding and exceptions.

Binary session types guarantee communication safety and session fidelity, but alone they cannot rule out deadlocks arising from the interleaving of different sessions. In Classical Processes (CP)$-$a process calculus based on classical linear logic$-$deadlock freedom is guaranteed by combining channel creation and parallel composition under the same logical cut rule. Similarly, in Good Variation (GV)$-$a linear concurrent $\lambda$-calculus$-$deadlock freedom is guaranteed by combining channel creation and thread spawning under the same operation, called fork. In both CP and GV, deadlock freedom is achieved at the expense of expressivity, as the only processes allowed are tree-structured. Dardha and Gay define Priority CP (PCP), which allows cyclic-structured processes and restores deadlock freedom by using priorities, in line with Kobayashi and Padovani. Following PCP, we present Priority GV (PGV), a variant of GV which decouples channel creation from thread spawning. Consequently, we type cyclic-structured processes and restore deadlock freedom by using priorities. We show that our type system is sound by proving subject reduction and progress. We define an encoding from PCP to PGV and prove that the encoding preserves typing and is sound and complete with respect to the operational semantics.

Abstract Neural networks are very successful at detecting patterns in noisy data, and have become the technology of choice in many fields. However, their usefulness is hampered by their susceptibility to adversarial attacks . Recently, many methods for measuring and improving a network’s robustness to adversarial perturbations have been proposed, and this growing body of research has given rise to numerous explicit or implicit notions of robustness. Connections between these notions are often subtle, and a systematic comparison between them is missing in the literature. In this paper we begin addressing this gap, by setting up general principles for the empirical analysis and evaluation of a network’s robustness as a mathematical property—during the network’s training phase, its verification, and after its deployment. We then apply these principles and conduct a case study that showcases the practical benefits of our general approach.

This paper introduces Hypersequent GV (HGV), a modular and extensible core calculus for functional programming with session types that enjoys deadlock freedom, confluence, and strong normalisation. HGV exploits hyper-environments, which are collections of type environments, to ensure that structural congruence is type preserving. As a consequence we obtain a tight operational correspondence between HGV and HCP, a hypersequent-based process-calculus interpretation of classical linear logic. Our translations from HGV to HCP and vice-versa both preserve and reflect reduction. HGV scales smoothly to support Girard’s Mix rule, a crucial ingredient for channel forwarding and exceptions.

Priority Sesh is a library for session-typed communication in Linear Haskell which offers strong compile-time correctness guarantees. Priority Sesh offers two deadlock-free APIs for session-typed communication. The first guarantees deadlock freedom by restricting the process structure to trees and forests. It is simple and composeable, but rules out cyclic structures. The second guarantees deadlock freedom via priorities, which allows the programmer to safely use cyclic structures as well.

In this invited talk, we discuss state of the art in neural network verification. We propose the term continuous verification to characterise the family of methods that explore continuous nature of machine learning algorithms. We argue that methods of continuous verification must rely on robust programming language infrastructure (refinement types, automated proving, type-driven program synthesis), which provides a major opportunity for the declarative programming language community. Keywords: Neural Networks, Verification, AI.

We describe a design for generics in Go inspired by previous work on Featherweight Java by Igarashi, Pierce, and Wadler. Whereas subtyping in Java is nominal, in Go it is structural, and whereas generics in Java are defined via erasure, in Go we use monomorphisation. Although monomorphisation is widely used, we are one of the first to formalise it. Our design also supports a solution to The Expression Problem.

One of the leading textbooks for formal methods is Software Foundations (SF), written by Benjamin Pierce in collaboration with others, and based on Coq. After five years using SF in the classroom, we came to the conclusion that Coq is not the best vehicle for this purpose, as too much of the course needs to focus on learning tactics for proof derivation, to the cost of learning programming language theory. Accordingly, we have written a new textbook, Programming Language Foundations in Agda (PLFA). PLFA covers much of the same ground as SF, although it is not a slavish imitation. What did we learn from writing PLFA? First, that it is possible. One might expect that without proof tactics that the proofs become too long, but in fact proofs in PLFA are about the same length as those in SF. Proofs in Coq require an interactive environment to be understood, while proofs in Agda can be read on the page. Second, that constructive proofs of preservation and progress give immediate rise to a prototype evaluator. This fact is obvious in retrospect but it is not exploited in SF (which instead provides a separate normalise tactic) nor can we find it in the literature. Third, that using extrinsically-typed terms is far less perspicuous than using intrinsically-typed terms. SF uses the former presentation, while PLFA presents both; the former uses about 1.6 as many lines of Agda code as the latter, roughly the golden ratio. The textbook is written as a literate Agda script, and can be found here: http://plfa.inf.ed.ac.uk

We present Hypersequent Classical Processes (HCP), a revised interpretation of the “Proofs as Processes” correspondence between linear logic and the π-calculus initially proposed by Abramsky [1994], and later developed by Bellin and Scott [1994], Caires and Pfenning [2010], and Wadler [2014], among others. HCP mends the discrepancies between linear logic and the syntax and observable semantics of parallel composition in the π-calculus, by conservatively extending linear logic to hyperenvironments (collections of environments, inspired by the hypersequents by Avron [1991]). Separation of environments in hyperenvironments is internalised by ⊗ and corresponds to parallel process behaviour. Thanks to this property, for the first time we are able to extract a labelled transition system (lts) semantics from proof rewritings. Leveraging the information on parallelism at the level of types, we obtain a logical reconstruction of the delayed actions that Merro and Sangiorgi [2004] formulated to model non-blocking I/O in the π-calculus. We define a denotational semantics for processes based on Brzozowski derivatives, and uncover that non-interference in HCP corresponds to Fubini’s theorem of double antiderivation. Having an lts allows us to validate HCP using the standard toolbox of behavioural theory. We instantiate bisimilarity and barbed congruence for HCP, and obtain a full abstraction result: bisimilarity, denotational equivalence, and barbed congruence coincide.