Thesis title: Deep clustering: new methods for dealing with image data
Recent improvements in deep learning techniques show new clustering opportunities to deal with complex high-dimensional data. Data clustering is an unsupervised method that aims at classifying n data points in k groups while complying with the concepts of inner cohesion and external separation. Hard clustering provides that each data point belongs to only one cluster while fuzzy clustering provides that each data point may belong to multiple clusters with different membership degree values. To overcome the problems incurred by conventional clustering methods when dealing with complex data, a new technique called deep clustering has been developed. The main idea of deep clustering is to learn latent features of the input data using a deep neural network and, then, apply a clustering method to the resulting representation of the data. Two are the approaches of deep clustering: sequential clustering, which applies clustering on the learned deep neural network representation; and simultaneous clustering, which jointly optimizes the deep representation learning and clustering objective.
This work’s first step is to describe the main theoretical aspects underlying the proposed methods. We describe the area of artificial intelligence and machine learning and delve into the complexities of deep learning, clarifying the fundamental concepts behind neural networks. Then, the main concepts of clustering are described, showing the advantages and disadvantages of k-means and fuzzy k-means methods. Deep clustering is presented as a possible solution to overcome the problems of classical clustering methods; accordingly, we provide an overview of deep clustering with an insight into existing methods in the literature.
We present a new two-step clustering method based on the idea of ensemble clustering. Indeed, we combine the use of two different clustering methods: fuzzy k-means and k-means. The aim is to use the fuzzy method as a filter to find a restricted number of input data on which the crisp method can focus. Following this approach, we can improve the k-means performance through a less rigid method that can provide the complexity of the data classification. Moreover, we combine our proposal with a deep autoencoder neural network following the sequential deep clustering approach.
In the simultaneous deep clustering approach, we present a new method that combines a neural network with a fuzzy clustering method. We construct a method that links the encoder part of a deep autoencoder neural network to a new layer in which the membership degree values of the fuzzy k-means method are calculated. In order to make the procedure coherent with the fuzzy method, we jointly optimize the parameters by minimizing the fuzzy k-means objective function. In this scenario, we handle the possible distortion of the embedded space by adding a penalization term to the loss function.
To assess the effectiveness of the proposed methods on complex datasets, we applied them to image data. Empirical results revealed improvements in accuracy compared to classical clustering methods and several deep clustering methods already existing in the literature. These results underline the effectiveness of our proposals and place them as a valid alternative to existing methods. Furthermore, this work emphasizes the important role played by neural networks in overcoming the challenges posed by conventional clustering methodologies.