Detecting 2D and 3D clusters through feature extractions in deep convolutional neural networks

Abstract

Clustering data is a complex and computationally expensive task, and in some approaches, it requires multiple passes over the same data to ensure globally optimal results. Clustering is an unsupervised learning method, meaning the distinction between points is made solely from the dataset being explored, with no prior knowledge or examples from previously clustered data. Using previously analyzed work to guide future analysis is supervised learning, which is typically applied to other problem categories, such as classification.

Our research investigates the application of supervised learning techniques, specifically Convolutional Neural Networks (CNN), to perform clustering on projections of datasets in 2D and 3D visual representations using graphs and voxelization representations of data. These CNN models are designed for categorical output and can be used to guide the training process and leverage previously clustered data to learn representations in new, previously unseen datasets. However, that categorical output can only represent the number of clusters present. To extend this approach further, we explore extracting information from the CNN's processing layers to analyze the activation maps between the convolutional layers using our proposed SilhouetteGen algorithm to delineate cluster shapes and locations within the original input space. In later models, our algorithm also replaces the CNN's categorical output after training is complete to remove any restrictions on the prediction range.

Various benchmark datasets and cluster quality metrics are used to assess the feasibility of this approach relative to a widely used and well-researched clustering method. The primary goal of this analysis is to demonstrate the feasibility of deep CNN feature extraction for detecting cluster information in distance and density-based clustering problems without requiring individual point-wise distance calculations.