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.
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.
One of the key challenges in designing machine learning systems is to determine the right balance amongst several objectives, which also oftentimes are incommensurable and conflicting. For example, when designing deep neural networks (DNNs), one often has to trade-off between multiple objectives, such as accuracy, energy consumption, and inference time. Typically, there is no single configuration that performs equally well for all objectives. Consequently, one is interested in identifying Pareto-optimal designs. Although different multi-objective optimization algorithms have been developed to identify Pareto-optimal configurations, state-of-the-art multi-objective optimization methods do not consider the different evaluation costs attending the objectives under consideration. This is particularly important for optimizing DNNs: the cost arising on account of assessing the accuracy of DNNs is orders of magnitude higher than that of measuring the energy consumption of pre-trained DNNs. We propose FlexiBO, a flexible Bayesian optimization method, to address this issue. We formulate a new acquisition function based on the improvement of the Pareto hyper-volume weighted by the measurement cost of each objective. Our acquisition function selects the next sample and objective that provides maximum information gain per unit of cost. We evaluated FlexiBO on 7 state-of-the-art DNNs for object detection, natural language processing, and speech recognition. Our results indicate that, when compared to other state-of-the-art methods across the 7 architectures we tested, the Pareto front obtained using FlexiBO has, on average, a 28.44% higher contribution to the true Pareto front and achieves 25.64% better diversity.
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.
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.
This document provides and in-depth introduction to the mlr framework for machine learning experiments in R.
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.