Self-supervised autoencoders for clustering and classification

  • PDF / 2,564,573 Bytes
  • 14 Pages / 595.276 x 790.866 pts Page_size
  • 112 Downloads / 279 Views

DOWNLOAD

REPORT


ORIGINAL PAPER

Self‑supervised autoencoders for clustering and classification Paraskevi Nousi1   · Anastasios Tefas1 Received: 4 January 2018 / Accepted: 13 May 2018 © Springer-Verlag GmbH Germany, part of Springer Nature 2018

Abstract Clustering techniques aim at finding meaningful groups of data samples which exhibit similarity with regards to a set of characteristics, typically measured in terms of pairwise distances. Due to the so-called curse of dimensionality, i.e., the observation that high-dimensional spaces are unsuited for measuring distances, distance-based clustering techniques such as the classic k-means algorithm fail to uncover meaningful clusters in high-dimensional spaces. Thus, dimensionality reduction techniques can be used to greatly improve the performance of such clustering methods. In this work, we study Autoencoders as Deep Learning tools for dimensionality reduction, and combine them with k-means clustering to learn low-dimensional representations which improve the clustering performance by enhancing intra-cluster relationships and suppressing intercluster ones, in a self-supervised manner. In the supervised paradigm, distance-based classifiers may also greatly benefit from robust dimensionality reduction techniques. The proposed method is evaluated via multiple experiments on datasets of handwritten digits, various objects and faces, and is shown to improve external cluster quality measuring criteria. A fully supervised counterpart is also evaluated on two face recognition datasets, and is shown to improve the performance of various lightweight classifiers, allowing their use in real-time applications on devices with limited computational resources, such as Unmanned Aerial Vehicles (UAVs). Keywords  Autoencoders · Clustering · Classification · Dimensionality reduction · Deep learning

1 Introduction Clustering refers to the process of identifying groups of samples from a data set, which exhibit similarity with regards to a set of features (Chrysouli and Tefas 2015; Passalis and Tefas 2016). Clustering techniques are typically fully unsupervised, in the sense that they don’t make use of any prior knowledge on the data, such as labels which might accompany the data points. This distinguishes the clustering task from classification, where the prior knowledge of class labels of known data points is exploited for the purpose of generalizing and making predictions for unfamiliar samples. Amongst the most widely used clustering techniques, k-means (MacQueen et al. 1967) is perhaps the most popular one, although many different clustering algorithms have * Paraskevi Nousi [email protected] Anastasios Tefas [email protected] 1



Department of Informatics, Aristotle University of Thessaloniki, 54124 Thessaloniki, Greece

been proposed since its first introduction, many of which are summarized in Jain (2010). Despite the plethora of algorithms that have been proposed for the purpose of clustering, the popularity of k-means lies in its simplicity as well as its versatility and extensibility