Unsupervised and Semi-supervised Learning Methods based on Deep Clustering and Explainable Active Learning

Abstract

In the recent years, the progress of deep learning in terms of performance in various areas has enabled the extensive use of big data. More and more intelligent systems, either in industrial or medical areas, are based on the information extracted by machine learning algorithms from thousands of data. As deep learning based methods require an excessively large amount of data in order to be efficient, we can expect that the number of data needed in order to build intelligent systems will only continue to increase exponentially. This situation poses a challenge in terms of data labeling. In fact, most of the deep learning applications, especially in the medical area, are based on the supervised learning. Even though the supervised learning paradigm reaches tremendous results in terms of discrimination, the necessity of having thousands of labeled images represents a serious drawback for this kind of learning system. The aim of the present study is to explore different methods, from deep clustering to semi-supervised learning techniques, that can completely eliminate or substantially diminish the need of data labeling process while maintaining a high performance in terms of discrimination. Our study focuses on the classification of the different types of the Human Epithelial type-2 (HEp-2) cells. We propose two models that are based on different scenarios. The first model is based on an unsupervised learning paradigm (scheme). We can say that the method is fully unsupervised in the sense that there is no need of labeled data in order to perform an end-to-end training of the model. On that purpose, we use deep clustering, a technique that combines conventional clustering methods with deep learning structures. A clustering layer is incorporated in the middle of a deep convolutional autoencoder (DCAE) in order to perform clustering on the latent space's features at every iteration of the training process. This technique allows the DCAE to produce latent features that are easily discriminable. Furtherly, unlike in the actual state-of-the-art deep clustering methods, where the reconstruction of the DCAE is not maximized, we propose a novel architecture for deep clustering, named attention-based deep clustering, where the DACE's reconstruction accuracy is efficiently maximized. Finally, for this first model, we present a systematic analysis of the results in order to explore the influence of the reconstruction accuracy on the latent features. The second model is based on a semi-supervised learning scheme. We adopt the techniques of active learning in order to alleviate the data labeling process. Active learning consists of applying some methods in order to select, among the thousands of available data, only those that really need to be labeled. It consists of selecting a limited number of informative data that can help to maximize the overall model's accuracy without the exigency of labeling the totality of the available data. In this work, we redefine active learning by proposing a data selection process based on explainable artificial intelligence (XAI). Relevance maps produced by some pre-selected XAI methods are used in order to find the informative data that need to be labeled. This proposed active learning scheme is performance-agnostic, since the data selection process does not depend on the accuracy of the model like in the conventional active learning methods. Secondly, by using XAI, we ensure the explainability of the selection process, which can contribute to the overall understating of the model. The two proposed models, the deep clustering and the explainable active learning, are extensively tested on the existing HEp-2 cell public datasets. The key contributions of the present work can be summarized as follows. For the deep clustering model, (1) we propose different techniques (pooling indices storage, copy and concatenation and attention network) in order to boost the reconstruction of the DCAE. Our method outperforms the state-of-the-art deep clustering method in terms of reconstruction accuracy. (2) We investigate if the reconstruction quality can affect the clustering accuracy. (3) We demonstrate that a better reconstruction of the DCAE increases the quality of the learned latent features. Our best model (attention network) outperforms the conventional handcrafted features and reaches similar level (97.56% on one of the datasets) with the supervised learning methods in terms of discrimination of the HEp-2 cells. For the proposed explainable active learning, (1) we propose a new definition of the selection method. Instead of using the model's output, as normally done in the conventional methods, we propose the use of the relevance maps. (2) Following the obtained results during the experiments, this selection method provides better results and, more importantly, has the advantage of the explainability. This is very important in the context of interpretability issues posed by the deep learning models. (3) We demonstrate that, using our proposed active learning method, an XAI-based selection contributes to boost the performance of the model even with a quite limited number of labeled data. With only 20% of the training data, the discrimination results achieve 92.77% of accuracy in one of the large HEp-2 cell dataset, while random and conventional selections give, respectively, 82% and 90%.