Lab methods for machine learning

From ICT research methods
Jump to: navigation, search

Machine Learning (ML) software is a software system with one or more components that learn from data. This entails engineering a pipeline for the collection and pre-processing of data, the training of an ML model, the deployment of the trained model to perform inference and the software engineering of the encompassing software system that sends new input data to the model to get answers.

For Machine Learning projects, the existing Lab methods can be applied as follows:

  • A/B testing A good way to discover which version of the model delivers more value to the end user. You need to incorporate logging or diagnostics that helps you decide which model “works better”.
  • Component test ML systems are difficult to divide into components so you need to come up with a clever way of doing this. ML components typically also interface through the data pipeline or shared data, so be aware of any adverse effects of changing one components behaviour on other components that use its output data.
  • Computer simulation Simulation models can be used to provide predictions to the end user or to optimize processes. Simulation models (a digital twin) are also used in situations where it is difficult to collect data from physical systems. The digital twin is used to collect input data and to test how the system responds to predicted output data. The use of digital twins is typical for training reinforcement learning models. Note: for ML projects this is not a Lab method, but a Workshop method.
  • Data analytics For ML projects this is not merely a Lab method. Its components “data collection”, “apply ML algorithms”, “data validation” and “ML model validation” merit separate methods in the framework.
  • Hardware validation If your ML solution needs to run on hardware components this might be part of your project to validate the component first.
  • Non-functional test The [paper by Zhang et al.] provides a good overview of the properties to be tested for ML systems: correctness, overfitting degree, fairness, interpretability, robustness, security, data privacy, and efficiency. The paper also sums up a literature overview of the methods to test them. A practical translation of that can be found in this post by Petra Heck.
  • Security test Also for ML systems you need to find and prioritise vulnerabilities and determine their impact on the confidentiality, integrity and availability of information. Privacy issues are involved if the system processes personal data.
  • System test Since the system is self-learning (you did not program the rules yourself) it might be difficult to assess if the system is behaving as required. You need to think about how you will test this on beforehand, see Petra Heck's post on testing ML applications for pointers. Usually you will keep aside part of your data for the final testing.
  • Unit test See component test (above).
  • Usability testing It is also important for ML solutions to test how well the end user is supported in the task the user needs to perform. Make sure you have a proper front-end (UI) through which the user can interact with the ML model and its outcome.
  • Note: there are separate methods for testing data and model: Data quality check, Model validation and Model evaluation. See also this post on testing ML applications for pointers on these topics.