How do you test AI models in an agile way?

An Example I have seen while starting AI based application testing: 1. Fix your CI-CD pipeline 2. Focus on cloud implementation in your test architecture which include end to end test run on the same application protocols 2. Run the generative test output after every New build. 3. Always get the latest samples of application run in the product using MLflow and TenserBoard and improve your baseline test data continuously. 4. Always focus on accuracy, precision and error rate of your generated matrics and compare it with production baseline..

The most important aspect of developing an AI system is TESTING. The models need to be thoroughly tested for all possible scenarios, edge cases and biases. It is specially true for black-box models like deep learning based neural nets. As you can not interpret the decision making process of the model, you need to understand the results for different inputs. For these, its crucial to understand the business objective of developing a model and find a metric that aligns with it. You can have multiple metrics to evaluate overall model performance, but to compare results of 2 models and to make sure with your fine-tuning you are actually making progress, there needs to be only 1 North star metrics, that also aligns with business needs.

Testing AI models in an agile way involves a combination of iterative development, continuous integration, and feedback loops. The goal is to ensure that the AI model's functionality aligns with business requirements while also maintaining high-quality performance metrics. The important aspect is that the business case is aligned with the data and variables selected for the AI use case. From there the results need to be validated that be relevant and correct as expected for the use case.

First make sure the change being tested is really well defined. Was it adding support for a new variable? Tweaking weights? To test agile the quickest testing would be to exclude as many other factors as possible to only test the new functionality if possible. Alternatives would be testing pre and post change and monitoring for expected differences in the output metrics.

For testing AI models in an agile way you should: - Focus on iterative development. Break model building into small increments that can be tested frequently. Get feedback early and often. - Prioritize unit and integration testing. Unit tests validate individual model components work as expected. - Test models against representative datasets. - Monitor models closely after deploying. Have a rollback plan ready. - Collaborate closely with business teams. - Automate reporting where possible. Generate testing reports/analytics to track model quality over time. Share results frequently with stakeholders. The key is taking an iterative, collaborative approach to testing that gathers rapid feedback and allows continuously improving models.

Automation is essential for agile testing, as it can help to reduce manual effort, increase efficiency, improve consistency, and enable faster feedback loops. Automation can be applied to various aspects of testing AI models, such as data preparation, model training, model evaluation, model deployment, etc. However, automation does not mean that human involvement is not needed. Human testers still play a crucial role in designing, reviewing, analyzing, and interpreting the test results.

In healthcare, we sometimes use a ‘silent trial’ to see if the predicted behavior of the expert users matches what the model would recommend. For example let’s say you have a model that is predicting whether a contact center employee will call the patient upon reading a message from the patient portal You could deploy the model in production but not show the user the predictions. Then, after a period of time, you could compare the expert user’s actions with the model predictions and generate your accuracy metrics. You could keep refining the model, in an iterative way, until you are comfortable exposing the model to the users.

As the project evolves, automate as much of the testing and monitoring processes as possible. This includes automating data preparation, model training, testing, deployment, and monitoring. Scalable infrastructure and tools are essential for handling large-scale AI models and datasets.

AI models need to test both code and data for changes over time and impact on model results. Frameworks for both these aspects are required. Model and data drift should be detected and the impact on model results should be quantified at the point of update. Actions need to be taken if thresholds for accuracy are exceeded due to changes in either code or data before results reach production. I have used tools such as DVC, MLFlow, Weights and Biases and frameworks written in Base SAS to track model performance over time.

Software system = Code AI system = Code + Data + Model There is a difference between a conventional software system and an AI system. And so there should also be a difference in how its tested. AI needs to have the following tests: - Data testing (data drift, data validation, feature generation tests) - Model testing (output test, preprocessing test, confidence interval tests) - Code tests After each individual component testing, there should be integration tests between all the components.

To minimize issues during deployment, perform user acceptance testing (UAT) during the sprints. Unit testing is great, but can miss the mark if you are not building the right product. Need to build in Sprint UAT as part of the schedule and design. Additionally, this will help to understand how the users will respond once they get the solution. Building a product that the user does not like and/or cannot use as they expect is a missed testing level.

Take an iterative approach with continuous model validation at each sprint. Utilize test data sets to regularly assess quality of AI models. Automate this processing through automated testing for efficiency. Focus on incremental improvements rather than perfect performance. AI models get better through time, and close monitoring ensures that model improves in performance, rather than declines, over time.

This is not different to any other sensible test method when working in an agile Environment. The AI Model specific part of in this section is missing

It is important to test at all levels and stages of software development to ensure that the software is of high quality. Here are some tips for testing at different levels and stages Start testing early. Don't wait until the software system is complete to start testing Start testing as soon as you have written some code Test at different levels. Unit testing, integration testing, system testing, and acceptance testing all play an important role in ensuring the quality of the software system. Automate your tests. Automated tests can save you a lot of time and effort. Consider automating your unit tests and integration tests Involve users in testing. User testing is essential for ensuring that the software system meets the needs of its users.

Testing AI models comprehensively through various stages and aspects ensures early identification of discrepancies, enhancing overall model quality and aligning with stakeholder expectations. Such a meticulous approach not only mitigates risks but also fortifies the model's reliability and performance in real-world scenarios.

Be very careful with using synthetic data in testing models, as the underlying assumption is that the process that has generated the data must be imbuing the synthetic data with statistical properties resembling production data. Otherwise you are just training a dog detector by having it look at cats, and what's the point in that?

tips for using test data wisely: Use a test data management tool. A test data management tool can help you to generate, manage, and store your test data. Use a data masking tool. A data masking tool can help you to mask sensitive data in your test data. This can help you to protect your customers' privacy and security. Use a data anonymization tool. A data anonymization tool can help you to anonymize your test data. This can help you to protect your customers' privacy. Use a data validation tool. A data validation tool can help you to validate your test data to ensure that it is accurate and complete.

All the data that is used for AI, be it training or testing is very crucial. Not only quality, but quantity, similarity to each other and similarity to real world data is highly crucial. Especially for cases like imbalanced data or tasks like anomaly detection or for tasks like cancer detection where the data is limited, data splitting between train and test needs detailed thinking. Same is for synthetic data. You need to take into consideration of its correlation with actual data. Synthetic data should only be used for training models, never for testing. Even during training, it should be used in a combination of real data to maximize results.

Test data should represent the real-world data the AI model will encounter in production. Test data should also be diverse enough to cover different scenarios and edge cases that may affect the behaviour or outcome of the AI model. This should be updated regularly based on new data sources or feedback from previous testing iterations.

Data integrity is fundamental. The integrity of AI model testing hinges on utilizing diverse, relevant, and accurately labeled test data, ensuring models avoid pitfalls like overfitting and generalize well to unseen scenarios. Leveraging tools like Label Studio or Amazon SageMaker Ground Truth can streamline the annotation process, ensuring consistency and precision in test data preparation.

It helps to adhere and structure communication compliant with IBM's CRISP-DM methodology - or the more recent ASUM-DM. The methodologies are best described with a visual graph, which I can't put here, but think of it as a process that goes back and forth between business understanding, data understanding, data preparation, modelling, evaluation, and eventual deployment when evaluation meets criteria for solving the framed business problem.

Establish clear communication channels. Make sure that everyone involved in the testing process knows how to communicate with each other. This may involve using email, chat, video conferencing, or a combination of these tools. Hold regular meetings. Regular meetings can help to keep everyone on the same page and to identify and resolve any issues early on. Use a test management tool. A test management tool can help you track your tests, assign tests to team members, and communicate test results. Be open to feedback. Be open to feedback from other team members and stakeholders. This feedback can help you to improve your testing process and the quality of your software.

Agile testing for AI demands adaptability. Prioritize testing key functionalities and model performance metrics. Implement continuous integration and delivery (CI/CD) pipelines for rapid updates. Employ A/B testing to assess model changes. Maintain a test suite that covers edge cases and evolving requirements.

When designing AI models, it's critical to collaborate with subject matter experts from various disciplines to bridge the gap between technical expertise and domain knowledge. One should consider working with different roles and disciplines, such as developers, testers, domain experts, business analysts, etc. When done right, this helps to 1. Define clear requirements based on actual users 2. Design effective and realistic test cases and scenarios that can capture the complexity and variability of real-world situations 3. Analyze and interpret the test results and provide recommendations for improving the AI model. 4. Explain and justify the decisions of the AI model to the stakeholders and users, as well as to the wider industry.

Maintain clear documentation for your AI model, including model architecture, data sources, and testing procedures. This aids collaboration among team members.

Become an ISTQB Certified Tester - AI Testing (CT-AI) in order to be able to integrate the topic of AI into the fundamental testing process and speak the same language with colleagues.

Human oversight in testing AI models is crucial, particularly for identifying and mitigating biases that automated systems may overlook. While automated tests are efficient, they may not capture the nuances leading to biased outcomes. Human testers can provide a more thorough evaluation by scrutinizing the model's fairness and ethical alignment. They can also assess how the model performs in real-world scenarios, which automated tests might not fully replicate. This is especially important for AI models that have broad social and ethical implications. Therefore, combining automated efficiency with human judgment is essential for responsible AI testing.

With non-deterministic AI it is key to monitor the outputs of the models in Production on an ongoing basis. This is less about testing inline but more making sure that you can understand what the causes of problems were if they arise (because it is unlikely you will identify an issue before it happens). Non-deterministic AI will not always be correct and due to fine tuning and self learning there is always the risk that models regress rather than improve. Testing should be based on deterministic tests that can be used to identify changes on the non-deterministic AI to understand regression impacts.

Though agile development prioritizes Individuals and interactions over processes and tools, the use of project management tools / collaborative tools can not be overlooked during the agile testing of AI projects. A good tool that will aid 1. Collaboration and communication 2. Issue tracking and track the completion of user stories to ensure alignment with project goals, identified KPI, and established baselines; 3. Integrations with version control systems and continuous integration platforms enhance coordination of testing AI models in agile among team members. This helps in an agile way

Gain a deep understanding of the AI model, its architecture, features, and requirements. Collaboration between development, testing, and subject matter experts is crucial. Automate repetitive and deterministic tests, such as input-output validation or performance benchmarks. Use testing frameworks compatible with AI technologies. Integrate AI model testing into the CI/CD pipeline to automate testing whenever code changes are made. This ensures that tests are run frequently and consistently.