Identifying patterns in high-dimensional and complex data, such as images, requires techniques that extract meaningful features. Deep clustering combines the representation power of neural networks with classical clustering and has shown strong performance on such data. However, most approaches build on $k$-Means, inheriting its assumptions about cluster shapes, requiring the number of clusters to be specified in advance, and lacking an intuitive stopping criterion. We propose DeepSynC, the first synchronisation-based deep clustering algorithm that overcomes these limitations. It begins by identifying core points in the embedded space and assigning them to clusters. A novel cluster loss then synchronises similarly embedded objects, enabling the gradual assignment of further points. This combination of synchronisation-based loss and assignment strategy allows greater flexibility in cluster shape and introduces an automatic stopping condition for training.

Top