https://scripties.uba.uva.nl/search?id=723252

Data annotation for machine learning can be a time-consuming and inefficient process. Interactive machine learning aims to make this process more efficient. This study aims to determine in what way interactive machine learning can be used to classify visual art pieces into discrete classes. In this context, interactive machine learning is defined as a machine learning algorithm that incorporates user feedback into the training process. Furthermore, this study compares the performance of interactive machine learning to conventional computer vision methods. Interactive machine learning is expected to outperform conventional machine learning algorithms in terms of data efficiency.

To evaluate if interactive machine learning provides a viable computer vision approach, this method was applied to a subset of the open-source WikiArt dataset. Approximately two thousand images were annotated by a human expert to identify approximately 500 portrait and 500 landscape paintings. These images were subsequently transformed with a vision transformer that extracted the appropriate features of the images. The features extracted from these images were used with the interactive machine learning algorithm that employed a support vector machine for classification. Three variations of interactive machine learning were compared to find the best approach for deciding which images to present to the user. The results indicate that uncertainty sampling allows for the lowest number of images needing to be annotated with interactive machine learning.

The present findings suggest that interactive machine learning is a viable alternative to other machine learning methods, especially when data is not readily available and needs to be annotated by a human.