Daten und KI
The issue of data quality (and quantity) is a key topic in our projects at the beginning. Due to the focus on the technical focus of recipes, the question of quantity is often dominant, as the quality of the data tends to be good in the development environment. The starting point for every AI project is the selection of the appropriate AI modeling strategy. To do this, the task must be defined as precisely as possible. In the first steps, we focus on a few objectives, which we validate and implement in principle in QuickScan and then in the proof of concept. Thanks to our experience in this environment, we don’t have to experiment with evaluations, but instead take a targeted approach to the questions we are looking for. The selection of data from measurement, recipe or laboratory values is quick and targeted. We have already achieved good to very good results with 40 and over 10,000 data sets. One objective in the paint sector was to achieve a DeltaE of less than 1.5 across the entire color space. The myth that “a lot helps a lot” is widespread in the AI environment. In order to train an AI correctly, the quality of the training and test data is particularly important in addition to the quantity. This is where we use quality assurance procedures to detect inconsistencies in the data. In addition, topics such as overfitting and underfitting are addressed and we deliberately test the limits of our models. In summary, we are already significantly reducing complexity through our domain knowledge and are familiar with suitable AI modeling, which allows us to take a very focused approach to the topic of data for AI implementation.