Grouping works of art using deep embedded clustering
No Thumbnail Available
Date
Authors
Granger, Bryan
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Over the past decade, deep embedded clustering has been used to combine the encoding properties of deep learning along with clustering algorithms. After reducing a high-dimensional matrix to a latent representation, a clustering algorithm can be used to group similar feature vectors together. While this technique has been used on a wide range of data, it has excelled with popular training sets such as the MNIST and USPS datasets, which consist of single handwritten digits. In “Deep Convolutional Embedding for Digitized Painting Clustering,” Castellano and Vessio [2020] use a deep convolutional embedded clustering framework to cluster paintings. In this project, I analyze the effectiveness of this algorithm in creating logical clusters. By applying this computational framework to a dataset consisting of works from the collection of the Metropolitan Museum of Art’s Open Access Program, I explore both the visual connections between works as well as the paintings’ contextual metadata. This analysis also reveals the importance of a hyperparameter that balances the joint loss function for the model, which can affect the homogeneity of each cluster. In using this algorithm on an actual collection in a museum, this project assesses the applicability of using such a tool within art historical and curatorial processes.