In recent years, transformer networks [V aswani et al., 2017] have been established as a fundamental neural architecture powering state-of-the-art results in many applications, including language

In recent years, transformer networks [V aswani et al., 2017] have been established as a fundamental neural architecture powering state-of-the-art results in many applications, including language

Attention layers, as commonly used in transformers, form the backbone of modern deep learning, yet there is no mathematical description of their benefits and deficiencies as compared with other architectures. In this work we establish both positive and negative results on the representation power of attention layers, with a focus on intrinsic complexity parameters such as width, depth, and embedding dimension. On the positive side, we present a sparse averaging task, where recurrent networks and feedforward networks all have complexity scaling polynomially in the input size, whereas transformers scale merely logarithmically in the input size; furthermore, we use the same construction to show the necessity and role of a large embedding dimension in a transformer. On the negative side, we present a triple detection task, where attention layers in turn have complexity scaling linearly in the input size; as this scenario seems rare in practice, we also present natural variants that can be efficiently solved by attention layers. The proof techniques emphasize the value of communication complexity in the analysis of transformers and related models, and the role of sparse averaging as a prototypical attention task, which even finds use in the analysis of triple detection.

As a contribution to quantitative set-theoretic inferencing, a translation is proposed of conjunctions of literals of the forms x y \z, x y \ z, and z {x}, where x,y,z stand for variables ranging over the von Neumann universe of sets, into unquantified Boolean formulae of a rather simple conjunctive normal form. The formulae in the target language involve variables ranging over a Boolean ring of sets, along with a difference operator and relators designating equality, non-disjointness and inclusion. Moreover, the result of each translation is a conjunction of literals of the forms x y\z, x y\z and of implications whose antecedents are isolated literals and whose consequents are either inclusions (strict or non-strict) between variables, or equalities between variables. Besides reflecting a simple and natural semantics, which ensures satisfiability-preservation, the proposed translation has quadratic algorithmic time-complexity, and bridges two languages both of which are known to have an NP-complete satisfiability problem. Key words: Satisfiability problem, Computable set theory, Expressibility, Proof verification, NP-completeness, quantitative logical inference.

Belief revision games (BRGs) are concerned with the dynamics of the beliefs of a group of communicating agents. BRGs are "zero-player" games where at each step every agent revises her own beliefs by taking account for the beliefs of her acquaintances. Each agent is associated with a belief state defined on some finite propositional language. We provide a general definition for such games where each agent has her own revision policy, and show that the belief sequences of agents can always be finitely characterized. We then define a set of revision policies based on belief merging operators. We point out a set of appealing properties for BRGs and investigate the extent to which these properties are satisfied by the merging-based policies under consideration.

Answer Set Programming (ASP) is a well-established paradigm of declarative programming that has been developed in the field of logic programming and nonmonotonic reasoning. Advances in ASP solving technology are customarily assessed in competition events, as it happens for other closely-related problem-solving technologies like SAT/SMT, QBF, Planning and Scheduling. ASP Competitions are (usually) biennial events; however, the Fifth ASP Competition departs from tradition, in order to join the FLoC Olympic Games at the Vienna Summer of Logic 2014, which is expected to be the largest event in the history of logic. This edition of the ASP Competition series is jointly organized by the University of Calabria (Italy), the Aalto University (Finland), and the University of Genova (Italy), and is affiliated with the 30th International Conference on Logic Programming (ICLP 2014). It features a completely re-designed setup, with novelties involving the design of tracks, the scoring schema, and the adherence to a fixed modeling language in order to push the adoption of the ASP-Core-2 standard. Benchmark domains are taken from past editions, and best system packages submitted in 2013 are compared with new versions and solvers. To appear in Theory and Practice of Logic Programming (TPLP).