Reinforcement Learning (RL) is increasingly adopted to train agents that can deal with complex sequential tasks, such as driving an autonomous vehicle or controlling a humanoid robot. Correspondingly, novel approaches are needed to ensure that RL agents have been tested adequately before going to production. Among them, mutation testing is quite promising, especially under the assumption that the injected faults (mutations) mimic the real ones. In this paper, we first describe a taxonomy of real RL faults obtained by repository mining. Then, we present the mutation operators derived from such real faults and implemented in the tool$\mu \text{PRL}$. Finally, we discuss the experimental results, showing that$\mu \text{PRL}$is effective at discriminating strong from weak test generators, hence providing useful feedback to developers about the adequacy of the generated test scenarios.

With the increased popularity of Deep Neural Networks (DNNs), increases also the need for tools to assist developers in the DNN implementation, testing and debugging process. Several approaches have been proposed that automatically analyse and localise potential faults in DNNs under test. In this work, we evaluate and compare existing state-of-the-art fault localisation techniques, which operate based on both dynamic and static analysis of the DNN. The evaluation is performed on a benchmark consisting of both real faults obtained from bug reporting platforms and faulty models produced by a mutation tool. Our findings indicate that the usage of a single, specific ground truth (e.g. the human-defined one) for the evaluation of DNN fault localisation tools results in pretty low performance (maximum average recall of 0.33 and precision of 0.21). However, such figures increase when considering alternative, equivalent patches that exist for a given faulty DNN. The results indicate that DeepFD is the most effective tool, achieving an average recall of 0.55 and a precision of 0.37 on our benchmark.

The growing adoption of Deep Learning (DL) systems in all areas of life makes the exposure and repair of errors in such systems a task of paramount importance. Existing approaches for measuring test adequacy and producing automated repair patches for such systems rely heavily on the activation values of neurons. However, there is no empirical evidence that links the behaviour of a DL model in terms of its neuron activation patterns to the type of faults present in that DL model. In this work, we perform a large-scale empirical study in which we inject artificial faults into a DL model and check whether the same types of faults lead to similar activation patterns. To analyse the patterns, we propose the notion of the spectrum of a deep neural network (DNN), which is defined as the probability distribution of the activation values of the neurons of the DNN. We perform our analysis on 6 subject systems considering 24 different fault types, each associated with a specific DL mutation operator. Our results show that we can successfully identify the fault type based on the top 3 spectra similarity matches in 75% of the cases and clusters of similar spectra have low impurity (i.e., they contain mostly one or few fault types). To demonstrate a practical application of this finding, we train different classifiers to predict the error type in a DL model from the spectra of the activation values. Our results show that the prediction accuracy can be as high as 79%, when the classifier is trained on subject-specific spectra.

As Deep Neural Networks (DNNs) are rapidly being adopted within large software systems, software developers are increasingly required to design, train, and deploy such models into the systems they develop. Consequently, testing and improving the robustness of these models have received a lot of attention lately. However, relatively little effort has been made to address the difficulties developers experience when designing and training such models: if the evaluation of a model shows poor performance after the initial training, what should the developer change? We survey and evaluate existing state-of-the-art techniques that can be used to repair model performance, using a benchmark of both real-world mistakes developers made while designing DNN models and artificial faulty models generated by mutating the model code. The empirical evaluation shows that random baseline is comparable with or sometimes outperforms existing state-of-the-art techniques. However, for larger and more complicated models, all repair techniques fail to find fixes. Our findings call for further research to develop more sophisticated techniques for Deep Learning repair.

Deep Learning (DL) solutions are increasingly adopted, but how to test them remains a major open research problem. Existing and new testing techniques have been proposed for and adapted to DL systems, including mutation testing. However, no approach has investigated the possibility to simulate the effects of real DL faults by means of mutation operators. We have defined 35 DL mutation operators relying on 3 empirical studies about real faults in DL systems. We followed a systematic process to extract the mutation operators from the existing fault taxonomies, with a formal phase of conflict resolution in case of disagreement. We have implemented 24 of these DL mutation operators into DeepCrime, the first source-level pre-training mutation tool based on real DL faults. We have assessed our mutation operators to understand their characteristics: whether they produce interesting, i.e., killable but not trivial, mutations. Then, we have compared the sensitivity of our tool to the changes in the quality of test data with that of DeepMutation++, an existing post-training DL mutation tool.

Deep Learning (DL) components are routinely integrated into software systems that need to perform complex tasks such as image or natural language processing. The adequacy of the test data used to test such systems can be assessed by their ability to expose artificially injected faults (mutations) that simulate real DL faults.In this paper, we describe an approach to automatically generate new test inputs that can be used to augment the existing test set so that its capability to detect DL mutations increases. Our tool DeepMetis implements a search based input generation strategy. To account for the non-determinism of the training and the mutation processes, our fitness function involves multiple instances of the DL model under test. Experimental results show that DeepMetis is effective at augmenting the given test set, increasing its capability to detect mutants by 63% on average. A leave-one-out experiment shows that the augmented test set is capable of exposing unseen mutants, which simulate the occurrence of yet undetected faults.

The growing application of deep neural networks in safety-critical domains makes the analysis of faults that occur in such systems of enormous importance. In this paper we introduce a large taxonomy of faults in deep learning (DL) systems. We have manually analysed 1059 artefacts gathered from GitHub commits and issues of projects that use the most popular DL frameworks (TensorFlow, Keras and PyTorch) and from related Stack Overflow posts. Structured interviews with 20 researchers and practitioners describing the problems they have encountered in their experience have enriched our taxonomy with a variety of additional faults that did not emerge from the other two sources. Our final taxonomy was validated with a survey involving an additional set of 21 developers, confirming that almost all fault categories (13/15) were experienced by at least 50% of the survey participants.

Abstract Context: A Machine Learning based System (MLS) is a software system including one or more components that learn how to perform a task from a given data set. The increasing adoption of MLSs in safety critical domains such as autonomous driving, healthcare, and finance has fostered much attention towards the quality assurance of such systems. Despite the advances in software testing, MLSs bring novel and unprecedented challenges, since their behaviour is defined jointly by the code that implements them and the data used for training them. Objective: To identify the existing solutions for functional testing of MLSs, and classify them from three different perspectives: (1) the context of the problem they address, (2) their features, and (3) their empirical evaluation. To report demographic information about the ongoing research. To identify open challenges for future research. Method: We conducted a systematic mapping study about testing techniques for MLSs driven by 33 research questions. We followed existing guidelines when defining our research protocol so as to increase the repeatability and reliability of our results. Results: We identified 70 relevant primary studies, mostly published in the last years. We identified 11 problems addressed in the literature. We investigated multiple aspects of the testing approaches, such as the used/proposed adequacy criteria, the algorithms for test input generation, and the test oracles. Conclusions: The most active research areas in MLS testing address automated scenario/input generation and test oracle creation. MLS testing is a rapidly growing and developing research area, with many open challenges, such as the generation of realistic inputs and the definition of reliable evaluation metrics and benchmarks.