Hridesh Rajan

Sumon Biswas, Hridesh Rajan

Fairness of machine learning (ML) software has become a major concern in the recent past. Although recent research on testing and improving fairness have demonstrated impact on real-world software, providing fairness guarantee in practice is still lacking. Certification of ML models is challenging because of the complex decision-making process of the models. In this paper, we proposed Fairify, an SMT-based approach to verify individual fairness property in neural network (NN) models. Individual fairness ensures that any two similar individuals get similar treatment irrespective of their protected attributes e.g., race, sex, age. Verifying this fairness property is hard because of the global checking and non-linear computation nodes in NN. We proposed sound approach to make individual fairness verification tractable for the developers. The key idea is that many neurons in the NN always remain inactive when a smaller part of the input domain is considered. So, Fairify leverages whitebox access to the models in production and then apply formal analysis based pruning. Our approach adopts input partitioning and then prunes the NN for each partition to provide fairness certification or counterexample. We leveraged interval arithmetic and activation heuristic of the neurons to perform the pruning as necessary. We evaluated Fairify on 25 real-world neural networks collected from four different sources, and demonstrated the effectiveness, scalability and performance over baseline and closely related work. Fairify is also configurable based on the domain and size of the NN. Our novel formulation of the problem can answer targeted verification queries with relaxations and counterexamples, which have practical implications.

Usman Gohar, Sumon Biswas, Hridesh Rajan

Machine Learning (ML) software has been widely adopted in modern society, with reported fairness implications for minority groups based on race, sex, age, etc. Many recent works have proposed methods to measure and mitigate algorithmic bias in ML models. The existing approaches focus on single classifier-based ML models. However, real-world ML models are often composed of multiple independent or dependent learners in an ensemble (e.g., Random Forest), where the fairness composes in a non-trivial way. How does fairness compose in ensembles? What are the fairness impacts of the learners on the ultimate fairness of the ensemble? Can fair learners result in an unfair ensemble? Furthermore, studies have shown that hyperparameters influence the fairness of ML models. Ensemble hyperparameters are more complex since they affect how learners are combined in different categories of ensembles. Understanding the impact of ensemble hyperparameters on fairness will help programmers design fair ensembles. Today, we do not understand these fully for different ensemble algorithms. In this paper, we comprehensively study popular real-world ensembles: bagging, boosting, stacking and voting. We have developed a benchmark of 168 ensemble models collected from Kaggle on four popular fairness datasets. We use existing fairness metrics to understand the composition of fairness. Our results show that ensembles can be designed to be fairer without using mitigation techniques. We also identify the interplay between fairness composition and data characteristics to guide fair ensemble design. Finally, our benchmark can be leveraged for further research on fair ensembles. To the best of our knowledge, this is one of the first and largest studies on fairness composition in ensembles yet presented in the literature.

Abhiram Bellur, Mohammed Raihan Ullah, Fraol Batole et al.

The primary value of AI agents in software development lies in their ability to extend the developer's capacity for reasoning and action, not to supplant human involvement. To showcase how to use agents working in tandem with developers, we designed a novel approach for carrying out coordinated renaming. Coordinated renaming, where a single rename refactoring triggers refactorings in multiple, related identifiers, is a frequent yet challenging task. Developers must manually propagate these rename refactorings across numerous files and contexts, a process that is both tedious and highly error-prone. State-of-the-art heuristic-based approaches produce an overwhelming number of false positives, while vanilla Large Language Models (LLMs) provide incomplete suggestions due to their limited context and inability to interact with refactoring tools. This leaves developers with incomplete refactorings or burdens them with filtering too many false positives. Coordinated renaming is exactly the kind of repetitive task that agents can significantly reduce the developers' burden while keeping them in the driver's seat. We designed, implemented, and evaluated the first multi-agent framework that automates coordinated renaming. It operates on a key insight: a developer's initial refactoring is a clue to infer the scope of related refactorings. Our Scope Inference Agent first transforms this clue into an explicit, natural-language Declared Scope. The Planned Execution Agent then uses this as a strict plan to identify program elements that should undergo refactoring and safely executes the changes by invoking the IDE's own trusted refactoring APIs. Finally, the Replication Agent uses it to guide the project-wide search. We first conducted a formative study on the practice of coordinated renaming in 609K commits in 100 open-source projects and surveyed 205 developers ...

Giang Nguyen, Sumon Biswas, Hridesh Rajan

Machine learning (ML) is increasingly being used in critical decision-making software, but incidents have raised questions about the fairness of ML predictions. To address this issue, new tools and methods are needed to mitigate bias in ML-based software. Previous studies have proposed bias mitigation algorithms that only work in specific situations and often result in a loss of accuracy. Our proposed solution is a novel approach that utilizes automated machine learning (AutoML) techniques to mitigate bias. Our approach includes two key innovations: a novel optimization function and a fairness-aware search space. By improving the default optimization function of AutoML and incorporating fairness objectives, we are able to mitigate bias with little to no loss of accuracy. Additionally, we propose a fairness-aware search space pruning method for AutoML to reduce computational cost and repair time. Our approach, built on the state-of-the-art Auto-Sklearn tool, is designed to reduce bias in real-world scenarios. In order to demonstrate the effectiveness of our approach, we evaluated our approach on four fairness problems and 16 different ML models, and our results show a significant improvement over the baseline and existing bias mitigation techniques. Our approach, Fair-AutoML, successfully repaired 60 out of 64 buggy cases, while existing bias mitigation techniques only repaired up to 44 out of 64 cases.

Ali Ghanbari, Deepak-George Thomas, Muhammad Arbab Arshad et al.

Deep neural networks (DNNs) are susceptible to bugs, just like other types of software systems. A significant uptick in using DNN, and its applications in wide-ranging areas, including safety-critical systems, warrant extensive research on software engineering tools for improving the reliability of DNN-based systems. One such tool that has gained significant attention in the recent years is DNN fault localization. This paper revisits mutation-based fault localization in the context of DNN models and proposes a novel technique, named deepmufl, applicable to a wide range of DNN models. We have implemented deepmufl and have evaluated its effectiveness using 109 bugs obtained from StackOverflow. Our results show that deepmufl detects 53/109 of the bugs by ranking the buggy layer in top-1 position, outperforming state-of-the-art static and dynamic DNN fault localization systems that are also designed to target the class of bugs supported by deepmufl. Moreover, we observed that we can halve the fault localization time for a pre-trained model using mutation selection, yet losing only 7.55% of the bugs localized in top-1 position.

Sayem Mohammad Imtiaz, Fraol Batole, Astha Singh et al.

Can we take a recurrent neural network (RNN) trained to translate between languages and augment it to support a new natural language without retraining the model from scratch? Can we fix the faulty behavior of the RNN by replacing portions associated with the faulty behavior? Recent works on decomposing a fully connected neural network (FCNN) and convolutional neural network (CNN) into modules have shown the value of engineering deep models in this manner, which is standard in traditional SE but foreign for deep learning models. However, prior works focus on the image-based multiclass classification problems and cannot be applied to RNN due to (a) different layer structures, (b) loop structures, (c) different types of input-output architectures, and (d) usage of both nonlinear and logistic activation functions. In this work, we propose the first approach to decompose an RNN into modules. We study different types of RNNs, i.e., Vanilla, LSTM, and GRU. Further, we show how such RNN modules can be reused and replaced in various scenarios. We evaluate our approach against 5 canonical datasets (i.e., Math QA, Brown Corpus, Wiki-toxicity, Clinc OOS, and Tatoeba) and 4 model variants for each dataset. We found that decomposing a trained model has a small cost (Accuracy: -0.6%, BLEU score: +0.10%). Also, the decomposed modules can be reused and replaced without needing to retrain.

Samantha Syeda Khairunnesa, Shibbir Ahmed, Sayem Mohammad Imtiaz et al.

Recent work has shown that Machine Learning (ML) programs are error-prone and called for contracts for ML code. Contracts, as in the design by contract methodology, help document APIs and aid API users in writing correct code. The question is: what kinds of contracts would provide the most help to API users? We are especially interested in what kinds of contracts help API users catch errors at earlier stages in the ML pipeline. We describe an empirical study of posts on Stack Overflow of the four most often-discussed ML libraries: TensorFlow, Scikit-learn, Keras, and PyTorch. For these libraries, our study extracted 413 informal (English) API specifications. We used these specifications to understand the following questions. What are the root causes and effects behind ML contract violations? Are there common patterns of ML contract violations? When does understanding ML contracts require an advanced level of ML software expertise? Could checking contracts at the API level help detect the violations in early ML pipeline stages? Our key findings are that the most commonly needed contracts for ML APIs are either checking constraints on single arguments of an API or on the order of API calls. The software engineering community could employ existing contract mining approaches to mine these contracts to promote an increased understanding of ML APIs. We also noted a need to combine behavioral and temporal contract mining approaches. We report on categories of required ML contracts, which may help designers of contract languages.

Abhiram Bellur, Fraol Batole, Mohammed Raihan Ullah et al.

MOVEMETHOD is a hallmark refactoring. Despite a plethora of research tools that recommend which methods to move and where, these recommendations do not align with how expert developers perform MOVEMETHOD. Given the extensive training of Large Language Models and their reliance upon naturalness of code, they should expertly recommend which methods are misplaced in a given class and which classes are better hosts. Our formative study of 2016 LLM recommendations revealed that LLMs give expert suggestions, yet they are unreliable: up to 80% of the suggestions are hallucinations. We introduce the first LLM fully powered assistant for MOVEMETHOD refactoring that automates its whole end-to-end lifecycle, from recommendation to execution. We designed novel solutions that automatically filter LLM hallucinations using static analysis from IDEs and a novel workflow that requires LLMs to be self-consistent, critique, and rank refactoring suggestions. As MOVEMETHOD refactoring requires global, projectlevel reasoning, we solved the limited context size of LLMs by employing refactoring-aware retrieval augment generation (RAG). Our approach, MM-assist, synergistically combines the strengths of the LLM, IDE, static analysis, and semantic relevance. In our thorough, multi-methodology empirical evaluation, we compare MM-assist with the previous state-of-the-art approaches. MM-assist significantly outperforms them: (i) on a benchmark widely used by other researchers, our Recall@1 and Recall@3 show a 1.7x improvement; (ii) on a corpus of 210 recent refactorings from Open-source software, our Recall rates improve by at least 2.4x. Lastly, we conducted a user study with 30 experienced participants who used MM-assist to refactor their own code for one week. They rated 82.8% of MM-assist recommendations positively. This shows that MM-assist is both effective and useful.

Sayem Mohammad Imtiaz, Astha Singh, Fraol Batole et al.

Not a day goes by without hearing about the impressive feats of large language models (LLMs), and equally, not a day passes without hearing about their challenges. LLMs are notoriously vulnerable to biases in their dataset, leading to issues such as toxicity. While domain-adaptive training has been employed to mitigate these issues, these techniques often address all model parameters indiscriminately during the repair process, resulting in poor repair quality and reduced model versatility. In this paper, we introduce a novel dynamic slicing-based intent-aware LLM repair strategy, IRepair. This approach selectively targets the most error-prone sections of the model for repair. Specifically, we propose dynamically slicing the model's most sensitive layers that require immediate attention, concentrating repair efforts on those areas. This method enables more effective repairs with potentially less impact on the model's overall performance by altering a smaller portion of the model. We evaluated our technique on three models from the GPT2 and GPT-Neo families, with parameters ranging from 800M to 1.6B, in a toxicity mitigation setup. Our results show that IRepair repairs errors 43.6% more effectively while causing 46% less disruption to general performance compared to the closest baseline, direct preference optimization. Our empirical analysis also reveals that errors are more concentrated in a smaller section of the model, with the top 20% of layers exhibiting 773% more error density than the remaining 80\%. This highlights the need for selective repair. Additionally, we demonstrate that a dynamic selection approach is essential for addressing errors dispersed throughout the model, ensuring a robust and efficient repair.

Shibbir Ahmed, Hongyang Gao, Hridesh Rajan

Deep learning models are trained with certain assumptions about the data during the development stage and then used for prediction in the deployment stage. It is important to reason about the trustworthiness of the model's predictions with unseen data during deployment. Existing methods for specifying and verifying traditional software are insufficient for this task, as they cannot handle the complexity of DNN model architecture and expected outcomes. In this work, we propose a novel technique that uses rules derived from neural network computations to infer data preconditions for a DNN model to determine the trustworthiness of its predictions. Our approach, DeepInfer involves introducing a novel abstraction for a trained DNN model that enables weakest precondition reasoning using Dijkstra's Predicate Transformer Semantics. By deriving rules over the inductive type of neural network abstract representation, we can overcome the matrix dimensionality issues that arise from the backward non-linear computation from the output layer to the input layer. We utilize the weakest precondition computation using rules of each kind of activation function to compute layer-wise precondition from the given postcondition on the final output of a deep neural network. We extensively evaluated DeepInfer on 29 real-world DNN models using four different datasets collected from five different sources and demonstrated the utility, effectiveness, and performance improvement over closely related work. DeepInfer efficiently detects correct and incorrect predictions of high-accuracy models with high recall (0.98) and high F-1 score (0.84) and has significantly improved over prior technique, SelfChecker. The average runtime overhead of DeepInfer is low, 0.22 sec for all unseen datasets. We also compared runtime overhead using the same hardware settings and found that DeepInfer is 3.27 times faster than SelfChecker.

Mohammad Wardat, Breno Dantas Cruz, Wei Le et al.

Deep Neural Networks (DNNs) are used in a wide variety of applications. However, as in any software application, DNN-based apps are afflicted with bugs. Previous work observed that DNN bug fix patterns are different from traditional bug fix patterns. Furthermore, those buggy models are non-trivial to diagnose and fix due to inexplicit errors with several options to fix them. To support developers in locating and fixing bugs, we propose DeepDiagnosis, a novel debugging approach that localizes the faults, reports error symptoms and suggests fixes for DNN programs. In the first phase, our technique monitors a training model, periodically checking for eight types of error conditions. Then, in case of problems, it reports messages containing sufficient information to perform actionable repairs to the model. In the evaluation, we thoroughly examine 444 models -53 real-world from GitHub and Stack Overflow, and 391 curated by AUTOTRAINER. DeepDiagnosis provides superior accuracy when compared to UMLUAT and DeepLocalize. Our technique is faster than AUTOTRAINER for fault localization. The results show that our approach can support additional types of models, while state-of-the-art was only able to handle classification ones. Our technique was able to report bugs that do not manifest as numerical errors during training. Also, it can provide actionable insights for fix whereas DeepLocalize can only report faults that lead to numerical errors during training. DeepDiagnosis manifests the best capabilities of fault detection, bug localization, and symptoms identification when compared to other approaches.

Giang Nguyen, Md Johir Islam, Rangeet Pan et al.

Today deep learning is widely used for building software. A software engineering problem with deep learning is that finding an appropriate convolutional neural network (CNN) model for the task can be a challenge for developers. Recent work on AutoML, more precisely neural architecture search (NAS), embodied by tools like Auto-Keras aims to solve this problem by essentially viewing it as a search problem where the starting point is a default CNN model, and mutation of this CNN model allows exploration of the space of CNN models to find a CNN model that will work best for the problem. These works have had significant success in producing high-accuracy CNN models. There are two problems, however. First, NAS can be very costly, often taking several hours to complete. Second, CNN models produced by NAS can be very complex that makes it harder to understand them and costlier to train them. We propose a novel approach for NAS, where instead of starting from a default CNN model, the initial model is selected from a repository of models extracted from GitHub. The intuition being that developers solving a similar problem may have developed a better starting point compared to the default model. We also analyze common layer patterns of CNN models in the wild to understand changes that the developers make to improve their models. Our approach uses commonly occurring changes as mutation operators in NAS. We have extended Auto-Keras to implement our approach. Our evaluation using 8 top voted problems from Kaggle for tasks including image classification and image regression shows that given the same search time, without loss of accuracy, Manas produces models with 42.9% to 99.6% fewer number of parameters than Auto-Keras' models. Benchmarked on GPU, Manas' models train 30.3% to 641.6% faster than Auto-Keras' models.

Sumon Biswas, Mohammad Wardat, Hridesh Rajan

Increasingly larger number of software systems today are including data science components for descriptive, predictive, and prescriptive analytics. The collection of data science stages from acquisition, to cleaning/curation, to modeling, and so on are referred to as data science pipelines. To facilitate research and practice on data science pipelines, it is essential to understand their nature. What are the typical stages of a data science pipeline? How are they connected? Do the pipelines differ in the theoretical representations and that in the practice? Today we do not fully understand these architectural characteristics of data science pipelines. In this work, we present a three-pronged comprehensive study to answer this for the state-of-the-art, data science in-the-small, and data science in-the-large. Our study analyzes three datasets: a collection of 71 proposals for data science pipelines and related concepts in theory, a collection of over 105 implementations of curated data science pipelines from Kaggle competitions to understand data science in-the-small, and a collection of 21 mature data science projects from GitHub to understand data science in-the-large. Our study has led to three representations of data science pipelines that capture the essence of our subjects in theory, in-the-small, and in-the-large.

Tianxiang Gao, Hailiang Liu, Jia Liu et al.

Implicit deep learning has received increasing attention recently due to the fact that it generalizes the recursive prediction rules of many commonly used neural network architectures. Its prediction rule is provided implicitly based on the solution of an equilibrium equation. Although a line of recent empirical studies has demonstrated its superior performances, the theoretical understanding of implicit neural networks is limited. In general, the equilibrium equation may not be well-posed during the training. As a result, there is no guarantee that a vanilla (stochastic) gradient descent (SGD) training nonlinear implicit neural networks can converge. This paper fills the gap by analyzing the gradient flow of Rectified Linear Unit (ReLU) activated implicit neural networks. For an $m$-width implicit neural network with ReLU activation and $n$ training samples, we show that a randomly initialized gradient descent converges to a global minimum at a linear rate for the square loss function if the implicit neural network is \textit{over-parameterized}. It is worth noting that, unlike existing works on the convergence of (S)GD on finite-layer over-parameterized neural networks, our convergence results hold for implicit neural networks, where the number of layers is \textit{infinite}.

Rangeet Pan, Hridesh Rajan

Training from scratch is the most common way to build a Convolutional Neural Network (CNN) based model. What if we can build new CNN models by reusing parts from previously build CNN models? What if we can improve a CNN model by replacing (possibly faulty) parts with other parts? In both cases, instead of training, can we identify the part responsible for each output class (module) in the model(s) and reuse or replace only the desired output classes to build a model? Prior work has proposed decomposing dense-based networks into modules (one for each output class) to enable reusability and replaceability in various scenarios. However, this work is limited to the dense layers and based on the one-to-one relationship between the nodes in consecutive layers. Due to the shared architecture in the CNN model, prior work cannot be adapted directly. In this paper, we propose to decompose a CNN model used for image classification problems into modules for each output class. These modules can further be reused or replaced to build a new model. We have evaluated our approach with CIFAR-10, CIFAR-100, and ImageNet tiny datasets with three variations of ResNet models and found that enabling decomposition comes with a small cost (1.77% and 0.85% for top-1 and top-5 accuracy, respectively). Also, building a model by reusing or replacing modules can be done with a 2.3% and 0.5% average loss of accuracy. Furthermore, reusing and replacing these modules reduces CO2e emission by ~37 times compared to training the model from scratch.

Sumon Biswas, Hridesh Rajan

In recent years, many incidents have been reported where machine learning models exhibited discrimination among people based on race, sex, age, etc. Research has been conducted to measure and mitigate unfairness in machine learning models. For a machine learning task, it is a common practice to build a pipeline that includes an ordered set of data preprocessing stages followed by a classifier. However, most of the research on fairness has considered a single classifier based prediction task. What are the fairness impacts of the preprocessing stages in machine learning pipeline? Furthermore, studies showed that often the root cause of unfairness is ingrained in the data itself, rather than the model. But no research has been conducted to measure the unfairness caused by a specific transformation made in the data preprocessing stage. In this paper, we introduced the causal method of fairness to reason about the fairness impact of data preprocessing stages in ML pipeline. We leveraged existing metrics to define the fairness measures of the stages. Then we conducted a detailed fairness evaluation of the preprocessing stages in 37 pipelines collected from three different sources. Our results show that certain data transformers are causing the model to exhibit unfairness. We identified a number of fairness patterns in several categories of data transformers. Finally, we showed how the local fairness of a preprocessing stage composes in the global fairness of the pipeline. We used the fairness composition to choose appropriate downstream transformer that mitigates unfairness in the machine learning pipeline.

Mohammad Wardat, Wei Le, Hridesh Rajan

Deep neural networks (DNNs) are becoming an integral part of most software systems. Previous work has shown that DNNs have bugs. Unfortunately, existing debugging techniques do not support localizing DNN bugs because of the lack of understanding of model behaviors. The entire DNN model appears as a black box. To address these problems, we propose an approach that automatically determines whether the model is buggy or not, and identifies the root causes. Our key insight is that historic trends in values propagated between layers can be analyzed to identify faults, and localize faults. To that end, we first enable dynamic analysis of deep learning applications: by converting it into an imperative representation and alternatively using a callback mechanism. Both mechanisms allows us to insert probes that enable dynamic analysis over the traces produced by the DNN while it is being trained on the training data. We then conduct dynamic analysis over the traces to identify the faulty layer that causes the error. We propose an algorithm for identifying root causes by capturing any numerical error and monitoring the model during training and finding the relevance of every layer on the DNN outcome. We have collected a benchmark containing 40 buggy models and patches that contain real errors in deep learning applications from Stack Overflow and GitHub. Our benchmark can be used to evaluate automated debugging tools and repair techniques. We have evaluated our approach using this DNN bug-and-patch benchmark, and the results showed that our approach is much more effective than the existing debugging approach used in the state of the practice Keras library. For 34 out of 40 cases, our approach was able to detect faults whereas the best debugging approach provided by Keras detected 32 out of 40 faults. Our approach was able to localize 21 out of 40 bugs whereas Keras did not localize any faults.

Sumon Biswas, Hridesh Rajan

Machine learning models are increasingly being used in important decision-making software such as approving bank loans, recommending criminal sentencing, hiring employees, and so on. It is important to ensure the fairness of these models so that no discrimination is made based on protected attribute (e.g., race, sex, age) while decision making. Algorithms have been developed to measure unfairness and mitigate them to a certain extent. In this paper, we have focused on the empirical evaluation of fairness and mitigations on real-world machine learning models. We have created a benchmark of 40 top-rated models from Kaggle used for 5 different tasks, and then using a comprehensive set of fairness metrics, evaluated their fairness. Then, we have applied 7 mitigation techniques on these models and analyzed the fairness, mitigation results, and impacts on performance. We have found that some model optimization techniques result in inducing unfairness in the models. On the other hand, although there are some fairness control mechanisms in machine learning libraries, they are not documented. The mitigation algorithm also exhibit common patterns such as mitigation in the post-processing is often costly (in terms of performance) and mitigation in the pre-processing stage is preferred in most cases. We have also presented different trade-off choices of fairness mitigation decisions. Our study suggests future research directions to reduce the gap between theoretical fairness aware algorithms and the software engineering methods to leverage them in practice.

Ramanathan Ramu, Ganesha B Upadhyaya, Hoan Anh Nguyen et al.

Many data-driven software engineering tasks such as discovering programming patterns, mining API specifications, etc., perform source code analysis over control flow graphs (CFGs) at scale. Analyzing millions of CFGs can be expensive and performance of the analysis heavily depends on the underlying CFG traversal strategy. State-of-the-art analysis frameworks use a fixed traversal strategy. We argue that a single traversal strategy does not fit all kinds of analyses and CFGs and propose bespoke control flow analysis (BCFA). Given a control flow analysis (CFA) and a large number of CFGs, BCFA selects the most efficient traversal strategy for each CFG. BCFA extracts a set of properties of the CFA by analyzing the code of the CFA and combines it with properties of the CFG, such as branching factor and cyclicity, for selecting the optimal traversal strategy. We have implemented BCFA in Boa, and evaluated BCFA using a set of representative static analyses that mainly involve traversing CFGs and two large datasets containing 287 thousand and 162 million CFGs. Our results show that BCFA can speedup the large scale analyses by 1%-28%. Further, BCFA has low overheads; less than 0.2%, and low misprediction rate; less than 0.01%.

Md Johirul Islam, Rangeet Pan, Giang Nguyen et al.

Significant interest in applying Deep Neural Network (DNN) has fueled the need to support engineering of software that uses DNNs. Repairing software that uses DNNs is one such unmistakable SE need where automated tools could be beneficial; however, we do not fully understand challenges to repairing and patterns that are utilized when manually repairing DNNs. What challenges should automated repair tools address? What are the repair patterns whose automation could help developers? Which repair patterns should be assigned a higher priority for building automated bug repair tools? This work presents a comprehensive study of bug fix patterns to address these questions. We have studied 415 repairs from Stack overflow and 555 repairs from Github for five popular deep learning libraries Caffe, Keras, Tensorflow, Theano, and Torch to understand challenges in repairs and bug repair patterns. Our key findings reveal that DNN bug fix patterns are distinctive compared to traditional bug fix patterns; the most common bug fix patterns are fixing data dimension and neural network connectivity; DNN bug fixes have the potential to introduce adversarial vulnerabilities; DNN bug fixes frequently introduce new bugs; and DNN bug localization, reuse of trained model, and coping with frequent releases are major challenges faced by developers when fixing bugs. We also contribute a benchmark of 667 DNN (bug, repair) instances.

Md Johirul Islam, Hoan Anh Nguyen, Rangeet Pan et al.

Modern software systems are increasingly including machine learning (ML) as an integral component. However, we do not yet understand the difficulties faced by software developers when learning about ML libraries and using them within their systems. To that end, this work reports on a detailed (manual) examination of 3,243 highly-rated Q&A posts related to ten ML libraries, namely Tensorflow, Keras, scikit-learn, Weka, Caffe, Theano, MLlib, Torch, Mahout, and H2O, on Stack Overflow, a popular online technical Q&A forum. We classify these questions into seven typical stages of an ML pipeline to understand the correlation between the library and the stage. Then we study the questions and perform statistical analysis to explore the answer to four research objectives (finding the most difficult stage, understanding the nature of problems, nature of libraries and studying whether the difficulties stayed consistent over time). Our findings reveal the urgent need for software engineering (SE) research in this area. Both static and dynamic analyses are mostly absent and badly needed to help developers find errors earlier. While there has been some early research on debugging, much more work is needed. API misuses are prevalent and API design improvements are sorely needed. Last and somewhat surprisingly, a tug of war between providing higher levels of abstractions and the need to understand the behavior of the trained model is prevalent.

Md Johirul Islam, Giang Nguyen, Rangeet Pan et al.

Deep learning has gained substantial popularity in recent years. Developers mainly rely on libraries and tools to add deep learning capabilities to their software. What kinds of bugs are frequently found in such software? What are the root causes of such bugs? What impacts do such bugs have? Which stages of deep learning pipeline are more bug prone? Are there any antipatterns? Understanding such characteristics of bugs in deep learning software has the potential to foster the development of better deep learning platforms, debugging mechanisms, development practices, and encourage the development of analysis and verification frameworks. Therefore, we study 2716 high-quality posts from Stack Overflow and 500 bug fix commits from Github about five popular deep learning libraries Caffe, Keras, Tensorflow, Theano, and Torch to understand the types of bugs, root causes of bugs, impacts of bugs, bug-prone stage of deep learning pipeline as well as whether there are some common antipatterns found in this buggy software. The key findings of our study include: data bug and logic bug are the most severe bug types in deep learning software appearing more than 48% of the times, major root causes of these bugs are Incorrect Model Parameter (IPS) and Structural Inefficiency (SI) showing up more than 43% of the times. We have also found that the bugs in the usage of deep learning libraries have some common antipatterns that lead to a strong correlation of bug types among the libraries.

Rangeet Pan, Md Johirul Islam, Shibbir Ahmed et al.

Despite numerous attempts to defend deep learning based image classifiers, they remain susceptible to the adversarial attacks. This paper proposes a technique to identify susceptible classes, those classes that are more easily subverted. To identify the susceptible classes we use distance-based measures and apply them on a trained model. Based on the distance among original classes, we create mapping among original classes and adversarial classes that helps to reduce the randomness of a model to a significant amount in an adversarial setting. We analyze the high dimensional geometry among the feature classes and identify the k most susceptible target classes in an adversarial attack. We conduct experiments using MNIST, Fashion MNIST, CIFAR-10 (ImageNet and ResNet-32) datasets. Finally, we evaluate our techniques in order to determine which distance-based measure works best and how the randomness of a model changes with perturbation.

John L. Singleton, Gary T. Leavens, Hridesh Rajan et al.

Modern software relies on libraries and uses them via application programming interfaces (APIs). Correct API usage as well as many software engineering tasks are enabled when APIs have formal specifications. In this work, we analyze the implementation of each method in an API to infer a formal postcondition. Conventional wisdom is that, if one has preconditions, then one can use the strongest postcondition predicate transformer (SP) to infer postconditions. However, SP yields postconditions that are exponentially large, which makes them difficult to use, either by humans or by tools. Our key idea is an algorithm that converts such exponentially large specifications into a form that is more concise and thus more usable. This is done by leveraging the structure of the specifications that result from the use of SP. We applied our technique to infer postconditions for over 2,300 methods in seven popular Java libraries. Our technique was able to infer specifications for 75.7% of these methods, each of which was verified using an Extended Static Checker. We also found that 84.6% of resulting specifications were less than 1/4 page (20 lines) in length. Our technique was able to reduce the length of SMT proofs needed for verifying implementations by 76.7% and reduced prover execution time by 26.7%.