Search

Three important criteria of existing convolutional neural networks (CNNs) are (1) test-set accuracy; (2) out-of-distribution accuracy; and (3) explainability. While these criteria have been studied independently, their relationship is unknown. For example, do CNNs that have a stronger out-of-distribution performance have also stronger explainability? Furthermore, most prior feature-importance studies only evaluate methods on 2-3 common vanilla ImageNet-trained CNNs, leaving it unknown how these methods generalize to CNNs of other architectures and training algorithms. Here, we perform the first, large-scale evaluation of the relations of the three criteria using 9 feature-importance methods and 12 ImageNet-trained CNNs that are of 3 training algorithms and 5 CNN architectures. We find several important insights and recommendations for ML practitioners. First, adversarially robust CNNs have a higher explainability score on gradient-based attribution methods (but not CAM-based or perturbation-based methods). Second, AdvProp models, despite being highly accurate more than both vanilla and robust models alone, are not superior in explainability. Third, among 9 feature attribution methods tested, GradCAM and RISE are consistently the best methods. Fourth, Insertion and Deletion are biased towards vanilla and robust models respectively, due to their strong correlation with the confidence score distributions of a CNN. Fifth, we did not find a single CNN to be the best in all three criteria, which interestingly suggests that CNNs are harder to interpret as they become more accurate.

Automated hyperparameter optimization (HPO) has gained great popularity and is an important ingredient of most automated machine learning frameworks. The process of designing HPO algorithms, however, is still an unsystematic and manual process: Limitations of prior work are identified and the improvements proposed are -- even though guided by expert knowledge -- still somewhat arbitrary. This rarely allows for gaining a holistic understanding of which algorithmic components are driving performance, and carries the risk of overlooking good algorithmic design choices. We present a principled approach to automated benchmark-driven algorithm design applied to multifidelity HPO (MF-HPO): First, we formalize a rich space of MF-HPO candidates that includes, but is not limited to common HPO algorithms, and then present a configurable framework covering this space. To find the best candidate automatically and systematically, we follow a programming-by-optimization approach and search over the space of algorithm candidates via Bayesian optimization. We challenge whether the found design choices are necessary or could be replaced by more naive and simpler ones by performing an ablation analysis. We observe that using a relatively simple configuration, in some ways simpler than established methods, performs very well as long as some critical configuration parameters have the right value., Comment: * Equal Contributions

Bayesian optimization is used in many areas of AI for the optimization of black-box processes and has achieved impressive improvements of the state of the art for a lot of applications. It intelligently explores large and complex design spaces while minimizing the number of evaluations of the expensive underlying process to be optimized. Materials science considers the problem of optimizing materials' properties given a large design space that defines how to synthesize or process them, with evaluations requiring expensive experiments or simulations -- a very similar setting. While Bayesian optimization is also a popular approach to tackle such problems, there is almost no overlap between the two communities that are investigating the same concepts. We present a survey of Bayesian optimization approaches in materials science to increase cross-fertilization and avoid duplication of work. We highlight common challenges and opportunities for joint research efforts.

A lot of technological advances depend on next-generation materials, such as graphene, which enables a raft of new applications, for example better electronics. Manufacturing such materials is often difficult; in particular, producing graphene at scale is an open problem. We provide a series of datasets that describe the optimization of the production of laser-induced graphene, an established manufacturing method that has shown great promise. We pose three challenges based on the datasets we provide -- modeling the behavior of laser-induced graphene production with respect to parameters of the production process, transferring models and knowledge between different precursor materials, and optimizing the outcome of the transformation over the space of possible production parameters. We present illustrative results, along with the code used to generate them, as a starting point for interested users. The data we provide represents an important real-world application of machine learning; to the best of our knowledge, no similar datasets are available.

The design of machine learning systems often requires trading off different objectives, for example, prediction error and energy consumption for deep neural networks (DNNs). Typically, no single design performs well in all objectives; therefore, finding Pareto-optimal designs is of interest. The search for Pareto-optimal designs involves evaluating designs in an iterative process, and the measurements are used to evaluate an acquisition function that guides the search process. However, measuring different objectives incurs different costs. For example, the cost of measuring the prediction error of DNNs is orders of magnitude higher than that of measuring the energy consumption of a pre-trained DNN, as it requires re-training the DNN. Current state-of-the-art methods do not consider this difference in objective evaluation cost, potentially incurring expensive evaluations of objective functions in the optimization process. In this paper, we develop a novel decoupled and cost-aware multi-objective optimization algorithm, we call Flexible Multi-Objective Bayesian Optimization (FlexiBO) to address this issue. FlexiBO weights the improvement of the hypervolume of the Pareto region by the measurement cost of each objective to balance the expense of collecting new information with the knowledge gained through objective evaluations, preventing us from performing expensive measurements for little to no gain. We evaluate FlexiBO on seven state-of-the-art DNNs for image recognition, natural language processing (NLP), and speech-to-text translation. Our results indicate that, given the same total experimental budget, FlexiBO discovers designs with 4.8$\%$ to 12.4$\%$ lower hypervolume error than the best method in state-of-the-art multi-objective optimization.

Modern deep neural network (DNN) systems are highly configurable with large a number of options that significantly affect their non-functional behavior, for example inference time and energy consumption. Performance models allow to understand and predict the effects of such configuration options on system behavior, but are costly to build because of large configuration spaces. Performance models from one environment cannot be transferred directly to another; usually models are rebuilt from scratch for different environments, for example different hardware. Recently, transfer learning methods have been applied to reuse knowledge from performance models trained in one environment in another. In this paper, we perform an empirical study to understand the effectiveness of different transfer learning strategies for building performance models of DNN systems. Our results show that transferring information on the most influential configuration options and their interactions is an effective way of reducing the cost to build performance models in new environments., Comment: 2 pages, 2 figures, USENIX Conference on Operational Machine Learning, 2019

The algorithm selection problem is to choose the most suitable algorithm for solving a given problem instance. It leverages the complementarity between different approaches that is present in many areas of AI. We report on the state of the art in algorithm selection, as defined by the Algorithm Selection competitions in 2015 and 2017. The results of these competitions show how the state of the art improved over the years. We show that although performance in some cases is very good, there is still room for improvement in other cases. Finally, we provide insights into why some scenarios are hard, and pose challenges to the community on how to advance the current state of the art.

Modern software systems in many application areas offer to the user a multitude of parameters, switches and other customisation hooks. Humans tend to have difficulties determining the best configurations for particular applications. Modern optimising compilers are an example of such software systems; their many parameters need to be tuned for optimal performance, but are often left at the default values for convenience. In this work, we automatically determine compiler parameter settings that result in optimised performance for particular applications. Specifically, we apply a state-of-the-art automated parameter configuration procedure based on cutting-edge machine learning and optimisation techniques to two prominent JavaScript compilers and demonstrate that significant performance improvements, more than 35% in some cases, can be achieved over the default parameter settings on a diverse set of benchmarks., Comment: 11 pages, long version of a poster presented at CGO 2016

This document provides and in-depth introduction to the mlr framework for machine learning experiments in R.

We present the results of the ICON Challenge on Algorithm Selection.

Constraint programming is used for a variety of real-world optimisation problems, such as planning, scheduling and resource allocation problems. At the same time, one continuously gathers vast amounts of data about these problems. Current constraint programming software does not exploit such data to update schedules, resources and plans. We propose a new framework, that we call the Inductive Constraint Programming loop. In this approach data is gathered and analyzed systematically, in order to dynamically revise and adapt constraints and optimization criteria. Inductive Constraint Programming aims at bridging the gap between the areas of data mining and machine learning on the one hand, and constraint programming on the other hand., Comment: 17 pages, 9 figures

The task of algorithm selection involves choosing an algorithm from a set of algorithms on a per-instance basis in order to exploit the varying performance of algorithms over a set of instances. The algorithm selection problem is attracting increasing attention from researchers and practitioners in AI. Years of fruitful applications in a number of domains have resulted in a large amount of data, but the community lacks a standard format or repository for this data. This situation makes it difficult to share and compare different approaches effectively, as is done in other, more established fields. It also unnecessarily hinders new researchers who want to work in this area. To address this problem, we introduce a standardized format for representing algorithm selection scenarios and a repository that contains a growing number of data sets from the literature. Our format has been designed to be able to express a wide variety of different scenarios. Demonstrating the breadth and power of our platform, we describe a set of example experiments that build and evaluate algorithm selection models through a common interface. The results display the potential of algorithm selection to achieve significant performance improvements across a broad range of problems and algorithms., Comment: Accepted to be published in Artificial Intelligence Journal

It is an unfortunate convention of science that research should pretend to be reproducible; our top tips will help you mitigate this fussy conventionality, enabling you to enthusiastically showcase your irreproducible work., Comment: 2 pages, LaTeX

This paper investigates the reproducibility of computational science research and identifies key challenges facing the community today. It is the result of the First Summer School on Experimental Methodology in Computational Science Research (https://blogs.cs.st-andrews.ac.uk/emcsr2014/). First, we consider how to reproduce experiments that involve human subjects, and in particular how to deal with different ethics requirements at different institutions. Second, we look at whether parallel and distributed computational experiments are more or less reproducible than serial ones. Third, we consider reproducible computational experiments from fields outside computer science. Our final case study looks at whether reproducibility for one researcher is the same as for another, by having an author attempt to have others reproduce their own, reproducible, paper. This paper is open, executable and reproducible: the whole process of writing this paper is captured in the source control repository hosting both the source of the paper, supplementary codes and data; we are providing setup for several experiments on which we were working; finally, we try to describe what we have achieved during the week of the school in a way that others may reproduce (and hopefully improve) our experiments., Comment: This paper was written at the First Summer School on Experimental Methodology in Computational Science Research, St Andrews, August 4-8, 2014, http://blogs.cs.st-andrews.ac.uk/emcsr2014

A common way of doing algorithm selection is to train a machine learning model and predict the best algorithm from a portfolio to solve a particular problem. While this method has been highly successful, choosing only a single algorithm has inherent limitations -- if the choice was bad, no remedial action can be taken and parallelism cannot be exploited, to name but a few problems. In this paper, we investigate how to predict the ranking of the portfolio algorithms on a particular problem. This information can be used to choose the single best algorithm, but also to allocate resources to the algorithms according to their rank. We evaluate a range of approaches to predict the ranking of a set of algorithms on a problem. We furthermore introduce a framework for categorizing ranking predictions that allows to judge the expressiveness of the predictive output. Our experimental evaluation demonstrates on a range of data sets from the literature that it is beneficial to consider the relationship between algorithms when predicting rankings. We furthermore show that relatively naive approaches deliver rankings of good quality already.

In recent years, portfolio approaches to solving SAT problems and CSPs have become increasingly common. There are also a number of different encodings for representing CSPs as SAT instances. In this paper, we leverage advances in both SAT and CSP solving to present a novel hierarchical portfolio-based approach to CSP solving, which we call Proteus, that does not rely purely on CSP solvers. Instead, it may decide that it is best to encode a CSP problem instance into SAT, selecting an appropriate encoding and a corresponding SAT solver. Our experimental evaluation used an instance of Proteus that involved four CSP solvers, three SAT encodings, and six SAT solvers, evaluated on the most challenging problem instances from the CSP solver competitions, involving global and intensional constraints. We show that significant performance improvements can be achieved by Proteus obtained by exploiting alternative view-points and solvers for combinatorial problem-solving., Comment: 11th International Conference on Integration of AI and OR Techniques in Constraint Programming for Combinatorial Optimization Problems. The final publication is available at link.springer.com

Algorithm portfolio and selection approaches have achieved remarkable improvements over single solvers. However, the implementation of such systems is often highly customised and specific to the problem domain. This makes it difficult for researchers to explore different techniques for their specific problems. We present LLAMA, a modular and extensible toolkit implemented as an R package that facilitates the exploration of a range of different portfolio techniques on any problem domain. It implements the algorithm selection approaches most commonly used in the literature and leverages the extensive library of machine learning algorithms and techniques in R. We describe the current capabilities and limitations of the toolkit and illustrate its usage on a set of example SAT problems.

The Algorithm Selection Problem is concerned with selecting the best algorithm to solve a given problem on a case-by-case basis. It has become especially relevant in the last decade, as researchers are increasingly investigating how to identify the most suitable existing algorithm for solving a problem instead of developing new algorithms. This survey presents an overview of this work focusing on the contributions made in the area of combinatorial search problems, where Algorithm Selection techniques have achieved significant performance improvements. We unify and organise the vast literature according to criteria that determine Algorithm Selection systems in practice. The comprehensive classification of approaches identifies and analyses the different directions from which Algorithm Selection has been approached. This paper contrasts and compares different methods for solving the problem as well as ways of using these solutions. It closes by identifying directions of current and future research.

Qualitative modelling is a technique integrating the fields of theoretical computer science, artificial intelligence and the physical and biological sciences. The aim is to be able to model the behaviour of systems without estimating parameter values and fixing the exact quantitative dynamics. Traditional applications are the study of the dynamics of physical and biological systems at a higher level of abstraction than that obtained by estimation of numerical parameter values for a fixed quantitative model. Qualitative modelling has been studied and implemented to varying degrees of sophistication in Petri nets, process calculi and constraint programming. In this paper we reflect on the strengths and weaknesses of existing frameworks, we demonstrate how recent advances in constraint programming can be leveraged to produce high quality qualitative models, and we describe the advances in theory and technology that would be needed to make constraint programming the best option for scientific investigation in the broadest sense., Comment: 15 pages plus references

We present a framework for a large-scale distributed eScience Artificial Intelligence search. Our approach is generic and can be used for many different problems. Unlike many other approaches, we do not require dedicated machines, homogeneous infrastructure or the ability to communicate between nodes. We give special consideration to the robustness of the framework, minimising the loss of effort even after total loss of infrastructure, and allowing easy verification of every step of the distribution process. In contrast to most eScience applications, the input data and specification of the problem is very small, being easily given in a paragraph of text. The unique challenges our framework tackles are related to the combinatorial explosion of the space that contains the possible solutions and the robustness of long-running computations. Not only is the time required to finish the computations unknown, but also the resource requirements may change during the course of the computation. We demonstrate the applicability of our framework by using it to solve a challenging and hitherto open problem in computational mathematics. The results demonstrate that our approach easily scales to computations of a size that would have been impossible to tackle in practice just a decade ago., Comment: 18 pages plus references. arXiv admin note: substantial text overlap with arXiv:1008.4328

We present preliminary results of an investigation into the suitability of virtualised hardware -- in particular clouds -- for running computational experiments. Our main concern was that the reported CPU time would not be reliable and reproducible. The results demonstrate that while this is true in cases where many virtual machines are running on the same physical hardware, there is no inherent variation introduced by using virtualised hardware compared to non-virtualised hardware.

Designing component-based constraint solvers is a complex problem. Some components are required, some are optional and there are interdependencies between the components. Because of this, previous approaches to solver design and modification have been ad-hoc and limited. We present a system that transforms a description of the components and the characteristics of the target constraint solver into a constraint problem. Solving this problem yields the description of a valid solver. Our approach represents a significant step towards the automated design and synthesis of constraint solvers that are specialised for individual constraint problem classes or instances.

Constraint problems can be trivially solved in parallel by exploring different branches of the search tree concurrently. Previous approaches have focused on implementing this functionality in the solver, more or less transparently to the user. We propose a new approach, which modifies the constraint model of the problem. An existing model is split into new models with added constraints that partition the search space. Optionally, additional constraints are imposed that rule out the search already done. The advantages of our approach are that it can be implemented easily, computations can be stopped and restarted, moved to different machines and indeed solved on machines which are not able to communicate with each other at all.

Constraint solvers are complex pieces of software which require many design decisions to be made by the implementer based on limited information. These decisions affect the performance of the finished solver significantly. Once a design decision has been made, it cannot easily be reversed, although a different decision may be more appropriate for a particular problem. We investigate using machine learning to make these decisions automatically depending on the problem to solve. We use the alldifferent constraint as a case study. Our system is capable of making non-trivial, multi-level decisions that improve over always making a default choice and can be implemented as part of a general-purpose constraint solver.

Programs to solve so-called constraint problems are complex pieces of software which require many design decisions to be made more or less arbitrarily by the implementer. These decisions affect the performance of the finished solver significantly. Once a design decision has been made, it cannot easily be reversed, although a different decision may be more appropriate for a particular problem. We investigate using machine learning to make these decisions automatically depending on the problem to solve with the alldifferent constraint as an example. Our system is capable of making non-trivial, multi-level decisions that improve over always making a default choice.

We report (to our knowledge) the first evaluation of Constraint Satisfaction as a computational framework for solving closest string problems. We show that careful consideration of symbol occurrences can provide search heuristics that provide several orders of magnitude speedup at and above the optimal distance. We also report (to our knowledge) the first analysis and evaluation -- using any technique -- of the computational difficulties involved in the identification of all closest strings for a given input set. We describe algorithms for web-scale distributed solution of closest string problems, both purely based on AI backtrack search and also hybrid numeric-AI methods.

This paper proposes a design for a system to generate constraint solvers that are specialised for specific problem models. It describes the design in detail and gives preliminary experimental results showing the feasibility and effectiveness of the approach.

This paper presents an evaluation of the design decisions made in four state-of-the-art constraint solvers; Choco, ECLiPSe, Gecode, and Minion. To assess the impact of design decisions, instances of the five problem classes n-Queens, Golomb Ruler, Magic Square, Social Golfers, and Balanced Incomplete Block Design are modelled and solved with each solver. The results of the experiments are not meant to give an indication of the performance of a solver, but rather investigate what influence the choice of algorithms and data structures has. The analysis of the impact of the design decisions focuses on the different ways of memory management, behaviour with increasing problem size, and specialised algorithms for specific types of variables. It also briefly considers other, less significant decisions.