FEATURE / AFP INSPECTION Figure 4. Prediction map and defect type color pallet. Figure 5. Operator defined defects. User Interface One of the notable trends in the literature, and what the authors consider a potential reason for the general resistance to the implementation of machine learning capabilities in physical systems, is a lack of comfort in the interaction with said ML systems. Thus, we have aimed to both improve operator relations to our software and alleviate some of the common industry concerns over ML applications. The ML inspection system outlined in this paper can be quickly understood and potentially corrected from an operator user interface (UI). Thus, an operator can react and correct system errors which can be recirculated for retraining of the network. This means that the ML algorithms implemented can be gradually and continuously improved 38 | SAMPE JOURNAL | through use. This directly addresses many of the common grievances against ML. Figures 3 to 5 display the mentioned functionalities on the UI. The main features of the USC developed UI include the following features, which the original UI provided by IMT included some variation of those features: M AY/J U N E 2 0 2 0 1. Display of defects with representative scan image 2. Operator management features and defect trackers 3. Operator correction capacity for misclassified defects 4. Capabilities for operator defined defects 5. Data management feature including export to AFP Defect Database and finite element export 6. Integration with the current IMT ACSIS automated inspection hardware w w w. s a m p e . o r ghttp://www.sampe.org