Recent advances in the verification of deep neural networks (DNNs) have opened the way for a broader usage of DNN verification technology in many application areas, including safety-critical ones. However, DNN verifiers are themselves complex programs that have been shown to be susceptible to errors and numerical imprecision; this, in turn, has raised the question of trust in DNN verifiers. One prominent attempt to address this issue is enhancing DNN verifiers with the capability of producing certificates of their results that are subject to independent algorithmic checking. While formulations of Marabou certificate checking already exist on top of the state-of-the-art DNN verifier Marabou, they are implemented in C++, and that code itself raises the question of trust (e.g., in the precision of floating point calculations or guarantees for implementation soundness). Here, we present an alternative implementation of the Marabou certificate checking in Imandra - an industrial functional programming language and an interactive theorem prover (ITP) - that allows us to obtain full proof of certificate correctness. The significance of the result is two-fold. Firstly, it gives stronger independent guarantees for Marabou proofs. Secondly, it opens the way for the wider adoption of DNN verifiers in interactive theorem proving in the same way as many ITPs already incorporate SMT solvers.

One can write dependently typed functional programs in Coq, and prove them correct in Coq; one can write low-level programs in C, and prove them correct with a C verification tool. We demonstrate how to write programs partly in Coq and partly in C, and interface the proofs together. The Verified Foreign Function Interface (VeriFFI) guarantees type safety and correctness of the combined program. It works by translating Coq function types (and constructor types) along with Coq functional models into VST function-specifications; if the user can prove in VST that the C functions satisfy those specs, then the C functions behave according to the user-specified functional models (even though the C implementation might be very different) and the proofs of Coq functions that call the C code can rely on that behavior. To achieve this translation, we employ a novel, hybrid deep/shallow description of Coq dependent types.

For performance and verification in machine learning, new methods have recently been proposed that optimise learning systems to satisfy formally expressed logical properties. Among these methods, differentiable logics (DLs) are used to translate propositional or first-order formulae into loss functions deployed for optimisation in machine learning. At the same time, recent attempts to give programming language support for verification of neural networks showed that DLs can be used to compile verification properties to machine-learning backends. This situation is calling for stronger guarantees about the soundness of such compilers, the soundness and compositionality of DLs, and the differentiability and performance of the resulting loss functions. In this paper, we propose an approach to formalise existing DLs using the Mathematical Components library in the Coq proof assistant. Thanks to this formalisation, we are able to give uniform semantics to otherwise disparate DLs, give formal proofs to existing informal arguments, find errors in previous work, and provide formal proofs to missing conjectured properties. This work is meant as a stepping stone for the development of programming language support for verification of machine learning.

Differentiable logics (DL) have recently been proposed as a method of training neural networks to satisfy logical specifications. A DL consists of a syntax in which specifications are stated and an interpretation function that translates expressions in the syntax into loss functions. These loss functions can then be used during training with standard gradient descent algorithms. The variety of existing DLs and the differing levels of formality with which they are treated makes a systematic comparative study of their properties and implementations difficult. This paper remedies this problem by suggesting a meta-language for defining DLs that we call the Logic of Differentiable Logics, or LDL. Syntactically, it generalises the syntax of existing DLs to FOL, and for the first time introduces the formalism for reasoning about vectors and learners. Semantically, it introduces a general interpretation function that can be instantiated to define loss functions arising from different existing DLs. We use LDL to establish several theoretical properties of existing DLs and to conduct their empirical study in neural network verification.

The mechanisation of the meta-theory of programming languages is still considered hard and requires considerable effort. When formalising properties of the extension of a language, one hence wants to reuse definitions and proofs. But type-theoretic proof assistants use inductive types and predicates to formalise syntax and type systems, and these definitions are closed to extensions. Available approaches for modular syntax are either inapplicable to type theory or add a layer of indirectness by requiring complicated encodings of types.

Call-by-push-value (CBPV) is an idealised calculus for functional and imperative programming, introduced as a subsuming paradigm for both call-by-value (CBV) and call-by-name (CBN). We formalise weak and strong operational semantics for (effect-free) CBPV, define its equational theory, and verify adequacy for the standard set/algebra denotational semantics. Furthermore, we prove normalisation of the standard reduction, confluence of strong reduction, strong normalisation using Kripke logical relations, and soundness of the equational theory using logical equivalence. We adapt and verify the known translations from CBV and CBN into CBPV for strong reduction. This yields, for instance, proofs of strong normalisation and confluence for the full λ-calculus with sums and products. Thanks to the automation provided by Coq and the Autosubst 2 framework, there is little formalisation overhead compared to detailed paper proofs.

Formalising metatheory in the Coq proof assistant is tedious as reasoning with binders without native support requires a lot of uninteresting technicalities. To relieve users from so-produced boilerplate, the Autosubst framework automates working with de Bruijn terms: For each annotated inductive type, Autosubst generates a corresponding instantiation operation for parallel substitutions and a decision procedure for assumption-free substitution lemmas. However, Autosubst is implemented in Ltac, Coq's tactic language, and thus suffers from Ltac's limitations. In particular, Autosubst is restricted to Coq and unscoped, non-mutual inductive types with a single sort of variables. In this paper, we present a new version of Autosubst that overcomes these restrictions. Autosubst 2 is an external code generator, which translates second-order HOAS specifications into potentially mutual inductive term sorts. We extend the equational theory of Autosubst to the case of mutual inductive sorts by combining the application of multiple parallel substitutions into exactly one instantiation operation for each sort, i.e. we parallelise substitutions to vector substitutions. The resulting equational theory is both simpler and more expressive than that of the original Autosubst framework and allows us to present an even more elegant proof of part A of the POPLMark challenge.

Abstract We propose a new collection of benchmark problems in mechanizing the metatheory of programming languages, in order to compare and push the state of the art of proof assistants. In particular, we focus on proofs using logical relations (LRs) and propose establishing strong normalization of a simply typed calculus with a proof by Kripke-style LRs as a benchmark. We give a modern view of this well-understood problem by formulating our LR on well-typed terms. Using this case study, we share some of the lessons learned tackling this problem in different dependently typed proof environments. In particular, we consider the mechanization in Beluga, a proof environment that supports higher-order abstract syntax encodings and contrast it to the development and strategies used in general-purpose proof assistants such as Coq and Agda. The goal of this paper is to engage the community in discussions on what support in proof environments is needed to truly bring mechanized metatheory to the masses and engage said community in the crafting of future benchmarks.

The de Bruijn representation of syntax with binding is commonly used, but flawed when it comes to recursion. As the structural recursion principle associated to an inductive type of expressions is unaware of the binding discipline, each recursive definition requires a separate proof of compatibility with variable instantiation. We solve this problem by extending Allais' notion of syntax traversals to obtain a framework for instantiation-compatible recursion. The framework is general enough to handle multivariate, potentially mutually recursive syntactic systems.