{"id":13187,"date":"2023-10-03T11:37:33","date_gmt":"2023-10-03T15:37:33","guid":{"rendered":"https:\/\/blogs.mathworks.com\/deep-learning\/?p=13187"},"modified":"2025-02-27T22:28:48","modified_gmt":"2025-02-28T03:28:48","slug":"verification-and-validation-for-ai-from-requirements-to-robust-modeling","status":"publish","type":"post","link":"https:\/\/blogs.mathworks.com\/deep-learning\/2023\/10\/03\/verification-and-validation-for-ai-from-requirements-to-robust-modeling\/","title":{"rendered":"Verification and Validation for AI: From requirements to robust modeling"},"content":{"rendered":"The following post is from <\/em>Lucas Garc\u00eda<\/em><\/a>, Product Manager for Deep Learning Toolbox.\u00a0<\/em>\r\n
<\/h6>\r\nThis is the second post in a 4-post series on Verification and Validation for AI. In the first post<\/a>, we gave an overview of the importance of Verification and Validation in AI within the context of AI Certification. We also introduced the W-shaped development process, an adaptation of the classical V-cycle to AI applications.\r\n
<\/h6>\r\n\"W-shaped\r\n
<\/h6>\r\nFigure 1:<\/strong> W-shaped development process. Credit: EASA, Daedalean.\r\n
<\/h6>\r\nThis blog post focuses on the steps you need to complete and iterate on starting with collecting requirements and up to creating a robust AI model. To illustrate the verification and validation steps from requirements to modeling, we will use as an example the design of a pneumonia detector.\r\n
<\/h6>\r\n \r\n
<\/h6>\r\n

Verification and Validation for Pneumonia Detection<\/strong><\/p>\r\nOur goal is to verify a deep learning model that identifies whether a patient is suffering from pneumonia by examining chest X-ray images. The image classification model needs to be not only accurate but also highly robust to avoid the potentially severe consequences of a misdiagnosis. We\u2019ll identify the problem and take it through all the steps in the W-shaped development process (W-cycle for short).\r\n

<\/h6>\r\nThe dataset we will be using is the MedMNISTv2 dataset. If you are familiar with MNIST for digit classification, MedMNIST<\/a> is a collection of labeled 2D and 3D biomedical lightweight 28 by 28 images. We decided to use this dataset because of its simplicity and the ability to rapidly iterate over the design. More specifically, we\u2019ll use the PneumoniaMNIST dataset, which is part of the MedMNISTv2 collection.\r\n
<\/h6>\r\n
\"MedMNIST<\/h6>\r\n
<\/h6>\r\nFigure 2:<\/strong> MedMNISTv2 dataset \u2013 The dataset is licensed under\u202fCreative Commons Attribution 4.0 International\u202f(CC BY 4.0<\/a>).\r\n
<\/h6>\r\nIn this post, we\u2019ll address the steps on the left-hand side of the W-cycle to create a Pneumonia Detector using the MedMNIST dataset, starting with Requirements allocated to ML component management<\/strong> all the way down to Model training<\/strong>. However, note that this is not a linear process, particularly when we evaluate the results of the training phase, so we\u2019ll have to iterate to refine our approach.\r\n
<\/h6>\r\n \r\n
<\/h6>\r\n

Requirements Allocated to ML Component Management<\/strong><\/p>\r\nWe\u2019ll start with the first step in the W-cycle related to AI and Machine Learning;<\/u> collecting the requirements specific to the Machine Learning component. Note that for any non-Machine Learning component items, you can follow the V-cycle frequently used for development assurance of traditional software.\r\n

<\/h6>\r\nAt this stage, key questions to consider are:\r\n