Scott Barnett

Tuan Dung Lai, Anj Simmons, Scott Barnett et al.

Background: Machine Learning (ML) systems rely on data to make predictions, the systems have many added components compared to traditional software systems such as the data processing pipeline, serving pipeline, and model training. Existing research on software maintenance has studied the issue-reporting needs and resolution process for different types of issues, such as performance and security issues. However, ML systems have specific classes of faults, and reporting ML issues requires domain-specific information. Because of the different characteristics between ML and traditional Software Engineering systems, we do not know to what extent the reporting needs are different, and to what extent these differences impact the issue resolution process. Objective: Our objective is to investigate whether there is a discrepancy in the distribution of resolution time between ML and non-ML issues and whether certain categories of ML issues require a longer time to resolve based on real issue reports in open-source applied ML projects. We further investigate the size of fix of ML issues and non-ML issues. Method: We extract issues reports, pull requests and code files in recent active applied ML projects from Github, and use an automatic approach to filter ML and non-ML issues. We manually label the issues using a known taxonomy of deep learning bugs. We measure the resolution time and size of fix of ML and non-ML issues on a controlled sample and compare the distributions for each category of issue.

Shangeetha Sivasothy, Scott Barnett, Stefanus Kurniawan et al.

Retrieval Augmented Generation (RAG) is increasingly being used when building Generative AI applications. Evaluating these applications and RAG pipelines is mostly done manually, via a trial and error process. Automating evaluation of RAG pipelines requires overcoming challenges such as context misunderstanding, wrong format, incorrect specificity, and missing content. Prior works therefore focused on improving evaluation metrics as well as enhancing components within the pipeline using available question and answer datasets. However, they have not focused on 1) providing a schema for capturing different types of question-answer pairs or 2) creating a set of templates for generating question-answer pairs that can support automation of RAG pipeline evaluation. In this paper, we present a technique for generating variations in question-answer pairs to trigger failures in RAG pipelines. We validate 5 open-source RAG pipelines using 3 datasets. Our approach revealed the highest failure rates when prompts combine multiple questions: 91% for questions when spanning multiple documents and 78% for questions from a single document; indicating a need for developers to prioritise handling these combined questions. 60% failure rate was observed in academic domain dataset and 53% and 62% failure rates were observed in open-domain datasets. Our automated approach outperforms the existing state-of-the-art methods, by increasing the failure rate by 51% on average per dataset. Our work presents an automated approach for continuously monitoring the health of RAG pipelines, which can be integrated into existing CI/CD pipelines, allowing for improved quality.

Jai Kannan, Scott Barnett, Luís Cruz et al.

Meeting the rise of industry demand to incorporate machine learning (ML) components into software systems requires interdisciplinary teams contributing to a shared code base. To maintain consistency, reduce defects and ensure maintainability, developers use code analysis tools to aid them in identifying defects and maintaining standards. With the inclusion of machine learning, tools must account for the cultural differences within the teams which manifests as multiple programming languages, and conflicting definitions and objectives. Existing tools fail to identify these cultural differences and are geared towards software engineering which reduces their adoption in ML projects. In our approach we attempt to resolve this problem by exploring the use of context which includes i) purpose of the source code, ii) technical domain, iii) problem domain, iv) team norms, v) operational environment, and vi) development lifecycle stage to provide contextualised error reporting for code analysis. To demonstrate our approach, we adapt Pylint as an example and apply a set of contextual transformations to the linting results based on the domain of individual project files under analysis. This allows for contextualised and meaningful error reporting for the end-user.

Dilani Widanapathiranage, Scott Barnett, Stefanus Kurniawan et al.

Hallucinations are a key concern when creating applications that rely on Foundation models (FMs). Understanding where and how these subtle failures occur in an application relies on evaluation methods known as \textit{evals}. Prior work focuses on defining new eval methods or benchmark datasets for specific tasks. However, neither helps a software team with a task-specific FM application when there is no metric or dataset. The demand for both automated approaches and deep integration of human insight makes this a challenging problem. We address this gap by proposing an approach to synthesise a FM task-specific evaluator program that provides automation and a custom UI for capturing feedback. The core novelty of our approach lies in: (1) a task-agnostic meta-model that captures properties of any FM task, (2) an interaction protocol for efficient use of human feedback, and (3) an eval synthesiser that selects or generates an appropriate set of evals. We implement our approach in \toolname and demonstrate the concept on two diverse FM tasks: chart data extraction and document question answering. A preliminary evaluation on the quality of our selected evals shows 93\% and 90\% accuracy respectively. Our research tackles a growing problem facing engineering teams, how to evaluate and review outputs from FM tasks.

Scott Barnett, Stefanus Kurniawan, Srikanth Thudumu et al.

Software engineers are increasingly adding semantic search capabilities to applications using a strategy known as Retrieval Augmented Generation (RAG). A RAG system involves finding documents that semantically match a query and then passing the documents to a large language model (LLM) such as ChatGPT to extract the right answer using an LLM. RAG systems aim to: a) reduce the problem of hallucinated responses from LLMs, b) link sources/references to generated responses, and c) remove the need for annotating documents with meta-data. However, RAG systems suffer from limitations inherent to information retrieval systems and from reliance on LLMs. In this paper, we present an experience report on the failure points of RAG systems from three case studies from separate domains: research, education, and biomedical. We share the lessons learned and present 7 failure points to consider when designing a RAG system. The two key takeaways arising from our work are: 1) validation of a RAG system is only feasible during operation, and 2) the robustness of a RAG system evolves rather than designed in at the start. We conclude with a list of potential research directions on RAG systems for the software engineering community.

Zafaryab Rasool, Scott Barnett, David Willie et al.

Large language models (LLMs) enable state-of-the-art semantic capabilities to be added to software systems such as semantic search of unstructured documents and text generation. However, these models are computationally expensive. At scale, the cost of serving thousands of users increases massively affecting also user experience. To address this problem, semantic caches are used to check for answers to similar queries (that may have been phrased differently) without hitting the LLM service. Due to the nature of these semantic cache techniques that rely on query embeddings, there is a high chance of errors impacting user confidence in the system. Adopting semantic cache techniques usually requires testing the effectiveness of a semantic cache (accurate cache hits and misses) which requires a labelled test set of similar queries and responses which is often unavailable. In this paper, we present VaryGen, an approach for using LLMs for test input generation that produces similar questions from unstructured text documents. Our novel approach uses the reasoning capabilities of LLMs to 1) adapt queries to the domain, 2) synthesise subtle variations to queries, and 3) evaluate the synthesised test dataset. We evaluated our approach in the domain of a student question and answer system by qualitatively analysing 100 generated queries and result pairs, and conducting an empirical case study with an open source semantic cache. Our results show that query pairs satisfy human expectations of similarity and our generated data demonstrates failure cases of a semantic cache. Additionally, we also evaluate our approach on Qasper dataset. This work is an important first step into test input generation for semantic applications and presents considerations for practitioners when calibrating a semantic cache.

Andrew J. Simmons, Scott Barnett, Jessica Rivera-Villicana et al.

Background: Meeting the growing industry demand for Data Science requires cross-disciplinary teams that can translate machine learning research into production-ready code. Software engineering teams value adherence to coding standards as an indication of code readability, maintainability, and developer expertise. However, there are no large-scale empirical studies of coding standards focused specifically on Data Science projects. Aims: This study investigates the extent to which Data Science projects follow code standards. In particular, which standards are followed, which are ignored, and how does this differ to traditional software projects? Method: We compare a corpus of 1048 Open-Source Data Science projects to a reference group of 1099 non-Data Science projects with a similar level of quality and maturity. Results: Data Science projects suffer from a significantly higher rate of functions that use an excessive numbers of parameters and local variables. Data Science projects also follow different variable naming conventions to non-Data Science projects. Conclusions: The differences indicate that Data Science codebases are distinct from traditional software codebases and do not follow traditional software engineering conventions. Our conjecture is that this may be because traditional software engineering conventions are inappropriate in the context of Data Science projects.

Anj Simmons, Scott Barnett, Anupam Chaudhuri et al.

Interpretable models are important, but what happens when the model is updated on new training data? We propose an algorithm for updating a decision tree while minimising the number of changes to the tree that a human would need to audit. We achieve this via a greedy approach that incorporates the number of changes to the tree as part of the objective function. We compare our algorithm to existing methods and show that it sits in a sweet spot between final accuracy and number of changes to audit.

Hala Abdelkader, Mohamed Abdelrazek, Scott Barnett et al.

Machine learning (ML), especially with the emergence of large language models (LLMs), has significantly transformed various industries. However, the transition from ML model prototyping to production use within software systems presents several challenges. These challenges primarily revolve around ensuring safety, security, and transparency, subsequently influencing the overall robustness and trustworthiness of ML models. In this paper, we introduce ML-On-Rails, a protocol designed to safeguard ML models, establish a well-defined endpoint interface for different ML tasks, and clear communication between ML providers and ML consumers (software engineers). ML-On-Rails enhances the robustness of ML models via incorporating detection capabilities to identify unique challenges specific to production ML. We evaluated the ML-On-Rails protocol through a real-world case study of the MoveReminder application. Through this evaluation, we emphasize the importance of safeguarding ML models in production.

Hira Naveed, Scott Barnett, Chetan Arora et al.

Context: Dynamic production environments make it challenging to maintain reliable machine learning (ML) systems. Runtime issues, such as changes in data patterns or operating contexts, that degrade model performance are a common occurrence in production settings. Monitoring enables early detection and mitigation of these runtime issues, helping maintain users' trust and prevent unwanted consequences for organizations. Aim: This study aims to provide a comprehensive overview of the ML monitoring literature. Method: We conducted a multivocal literature review (MLR) following the well established guidelines by Garousi to investigate various aspects of ML monitoring approaches in 136 papers. Results: We analyzed selected studies based on four key areas: (1) the motivations, goals, and context; (2) the monitored aspects, specific techniques, metrics, and tools; (3) the contributions and benefits; and (4) the current limitations. We also discuss several insights found in the studies, their implications, and recommendations for future research and practice. Conclusion: Our MLR identifies and summarizes ML monitoring practices and gaps, emphasizing similarities and disconnects between formal and gray literature. Our study is valuable for both academics and practitioners, as it helps select appropriate solutions, highlights limitations in current approaches, and provides future directions for research and tool development.

Srikanth Thudumu, Hy Nguyen, Hung Du et al.

Neural Architecture Search (NAS) aims to automate the design of deep neural networks. However, existing NAS techniques often focus on maximising accuracy, neglecting model efficiency. This limitation restricts their use in resource-constrained environments like mobile devices and edge computing systems. Moreover, current evaluation metrics prioritise performance over efficiency, lacking a balanced approach for assessing architectures suitable for constrained scenarios. To address these challenges, this paper introduces the M-factor, a novel metric combining model accuracy and size. Four diverse NAS techniques are compared: Policy-Based Reinforcement Learning, Regularised Evolution, Tree-structured Parzen Estimator (TPE), and Multi-trial Random Search. These techniques represent different NAS paradigms, providing a comprehensive evaluation of the M-factor. The study analyses ResNet configurations on the CIFAR-10 dataset, with a search space of 19,683 configurations. Experiments reveal that Policy-Based Reinforcement Learning and Regularised Evolution achieved M-factor values of 0.84 and 0.82, respectively, while Multi-trial Random Search attained 0.75, and TPE reached 0.67. Policy-Based Reinforcement Learning exhibited performance changes after 39 trials, while Regularised Evolution optimised within 20 trials. The research investigates the optimisation dynamics and trade-offs between accuracy and model size for each strategy. Findings indicate that, in some cases, random search performed comparably to more complex algorithms when assessed using the M-factor. These results highlight how the M-factor addresses the limitations of existing metrics by guiding NAS towards balanced architectures, offering valuable insights for selecting strategies in scenarios requiring both performance and efficiency.

Scott Barnett, Zac Brannelly, Stefanus Kurniawan et al.

Large Language Models (LLMs) have the unique capability to understand and generate human-like text from input queries. When fine-tuned, these models show enhanced performance on domain-specific queries. OpenAI highlights the process of fine-tuning, stating: "To fine-tune a model, you are required to provide at least 10 examples. We typically see clear improvements from fine-tuning on 50 to 100 training examples, but the right number varies greatly based on the exact use case." This study extends this concept to the integration of LLMs within Retrieval-Augmented Generation (RAG) pipelines, which aim to improve accuracy and relevance by leveraging external corpus data for information retrieval. However, RAG's promise of delivering optimal responses often falls short in complex query scenarios. This study aims to specifically examine the effects of fine-tuning LLMs on their ability to extract and integrate contextual data to enhance the performance of RAG systems across multiple domains. We evaluate the impact of fine-tuning on the LLMs' capacity for data extraction and contextual understanding by comparing the accuracy and completeness of fine-tuned models against baseline performances across datasets from multiple domains. Our findings indicate that fine-tuning resulted in a decline in performance compared to the baseline models, contrary to the improvements observed in standalone LLM applications as suggested by OpenAI. This study highlights the need for vigorous investigation and validation of fine-tuned models for domain-specific tasks.

Jai Kannan, Scott Barnett, Anj Simmons et al.

Deep learning models have become essential in software engineering, enabling intelligent features like image captioning and document generation. However, their popularity raises concerns about environmental impact and inefficient model selection. This paper introduces GreenRunnerGPT, a novel tool for efficiently selecting deep learning models based on specific use cases. It employs a large language model to suggest weights for quality indicators, optimizing resource utilization. The tool utilizes a multi-armed bandit framework to evaluate models against target datasets, considering tradeoffs. We demonstrate that GreenRunnerGPT is able to identify a model suited to a target use case without wasteful computations that would occur under a brute-force approach to model selection.

Alex Cummaudo, Scott Barnett, Rajesh Vasa et al.

Increased popularity of `intelligent' web services provides end-users with machine-learnt functionality at little effort to developers. However, these services require a decision threshold to be set which is dependent on problem-specific data. Developers lack a systematic approach for evaluating intelligent services and existing evaluation tools are predominantly targeted at data scientists for pre-development evaluation. This paper presents a workflow and supporting tool, Threshy, to help software developers select a decision threshold suited to their problem domain. Unlike existing tools, Threshy is designed to operate in multiple workflows including pre-development, pre-release, and support. Threshy is designed for tuning the confidence scores returned by intelligent web services and does not deal with hyper-parameter optimisation used in ML models. Additionally, it considers the financial impacts of false positives. Threshold configuration files exported by Threshy can be integrated into client applications and monitoring infrastructure. Demo: https://bit.ly/2YKeYhE.

Alex Cummaudo, Scott Barnett, Rajesh Vasa et al.

Intelligent services provide the power of AI to developers via simple RESTful API endpoints, abstracting away many complexities of machine learning. However, most of these intelligent services-such as computer vision-continually learn with time. When the internals within the abstracted 'black box' become hidden and evolve, pitfalls emerge in the robustness of applications that depend on these evolving services. Without adapting the way developers plan and construct projects reliant on intelligent services, significant gaps and risks result in both project planning and development. Therefore, how can software engineers best mitigate software evolution risk moving forward, thereby ensuring that their own applications maintain quality? Our proposal is an architectural tactic designed to improve intelligent service-dependent software robustness. The tactic involves creating an application-specific benchmark dataset baselined against an intelligent service, enabling evolutionary behaviour changes to be mitigated. A technical evaluation of our implementation of this architecture demonstrates how the tactic can identify 1,054 cases of substantial confidence evolution and 2,461 cases of substantial changes to response label sets using a dataset consisting of 331 images that evolve when sent to a service.

Maheswaree K Curumsing, Alex Cummaudo, Ulrike Maria Graetsch et al.

Software developers are increasingly using machine learning APIs to implement 'intelligent' features. Studies show that incorporating machine learning into an application increases technical debt, creates data dependencies, and introduces uncertainty due to non-deterministic behaviour. However, we know very little about the emotional state of software developers who deal with such issues. In this paper, we do a landscape analysis of emotion found in 1,245 Stack Overflow posts about computer vision APIs. We investigate the application of an existing emotion classifier EmoTxt and manually verify our results. We found that the emotion profile varies for different question categories.

Alex Cummaudo, Rajesh Vasa, Scott Barnett et al.

Intelligent services are becoming increasingly more pervasive; application developers want to leverage the latest advances in areas such as computer vision to provide new services and products to users, and large technology firms enable this via RESTful APIs. While such APIs promise an easy-to-integrate on-demand machine intelligence, their current design, documentation and developer interface hides much of the underlying machine learning techniques that power them. Such APIs look and feel like conventional APIs but abstract away data-driven probabilistic behaviour - the implications of a developer treating these APIs in the same way as other, traditional cloud services, such as cloud storage, is of concern. The objective of this study is to determine the various pain-points developers face when implementing systems that rely on the most mature of these intelligent services, specifically those that provide computer vision. We use Stack Overflow to mine indications of the frustrations that developers appear to face when using computer vision services, classifying their questions against two recent classification taxonomies (documentation-related and general questions). We find that, unlike mature fields like mobile development, there is a contrast in the types of questions asked by developers. These indicate a shallow understanding of the underlying technology that empower such systems. We discuss several implications of these findings via the lens of learning taxonomies to suggest how the software engineering community can improve these services and comment on the nature by which developers use them.

Andrew J. Simmons, Scott Barnett, Simon Vajda et al.

Data analysts often discover irregularities in their underlying dataset, which need to be traced back to the original source and corrected. Standards for representing data provenance (i.e. the origins of the data), such as the W3C PROV standard, can assist with this process, however require a mapping between abstract provenance concepts and the domain of use in order to apply them effectively. We propose a custom notation for expressing provenance of information in the sport performance analysis domain, and map our notation to concepts in the W3C PROV standard where possible. We evaluate the functionality of W3C PROV (without specialisations) and the VisTrails workflow manager (without extensions), and find that as is, neither are able to fully capture sport performance analysis workflows, notably due to limitations surrounding capture of automated and manual activities respectively. Furthermore, their notations suffer from ineffective use of visual design space, and present potential usability issues as their terminology is unlikely to match that of sport practitioners. Our findings suggest that one-size-fits-all provenance and workflow systems are a poor fit in practice, and that their notation and functionality need to be optimised for the domain of use.