Semi-supervised machine learning techniques for classification of evolving data in pattern recognition

摘要

The amount of data recorded and processed over recent years has increased exponentially. To create intelligent systems that can learn from this data, we need to be able to identify patterns hidden in the data itself, learn these pattern and predict future results based on our current observations. If we think about this system in the context of time, the data itself evolves and so does the nature of the classification problem. As more data become available, different classification algorithms are suitable for a particular setting. At the beginning of the learning cycle when we have a limited amount of data, online learning algorithms are more suitable. When truly large amounts of data become available, we need algorithms that can handle large amounts of data that might be only partially labeled as a result of the bottleneck in the learning pipeline from human labeling of the data.ududAn excellent example of evolving data is gesture recognition, and it is present throughout our work. We need a gesture recognition system to work fast and with very few examples at the beginning. Over time, we are able to collect more data and the system can improve. As the system evolves, the user expects it to work better and not to have to become involved when the classifier is unsure about decisions. This latter situation produces additional unlabeled data. Another example of an application is medical classification, where experts’ time is a rare resource and the amount of received and labeled data disproportionately increases over time.ududAlthough the process of data evolution is continuous, we identify three main discrete areas of contribution in different scenarios. When the system is very new and not enough data are available, online learning is used to learn after every single example and to capture the knowledge very fast. With increasing amounts of data, offline learning techniques are applicable. Once the amount of data is overwhelming and the teacher cannot provide labels for all the data, we have another setup that combines labeled and unlabeled data. These three setups define our areas of contribution; and our techniques contribute in each of them with applications to pattern recognition scenarios, such as gesture recognition and sketch recognition.ududAn online learning setup significantly restricts the range of techniques that can be used. In our case, the selected baseline technique is the Evolving TS-Fuzzy Model. The semi-supervised aspect we use is a relation between rules created by this model. Specifically, we propose a transductive similarity model that utilizes the relationship between generated rules based on their decisions about a query sample during the inference time. The activation of each of these rules is adjusted according to the transductive similarity, and the new decision is obtained using the adjusted activation. We also propose several new variations to the transductive similarity itself.ududOnce the amount of data increases, we are not limited to the online learning setup, and we can take advantage of the offline learning scenario, which normally performs better than the online one because of the independence of sample ordering and global optimization with respect to all samples. We use generative methods to obtain data outside of the training set. Specifically, we aim to improve the previously mentioned TS Fuzzy Model by incorporating semi-supervised learning in the offline learning setup without unlabeled data. We use the Universum learning approach and have developed a method called UFuzzy. This method relies on artificially generated examples with high uncertainty (Universum set), and it adjusts the cost function of the algorithm to force the decision boundary to be close to the Universum data. We were able to prove the hypothesis behind the design of the UFuzzy classifier that Universum learning can improve the TS Fuzzy Model and have achieved improved performance on more than two dozen datasets and applications.ududWith increasing amounts of data, we use the last scenario, in which the data comprises both labeled data and additional non-labeled data. This setting is one of the most common ones for semi-supervised learning problems. In this part of our work, we aim to improve the widely popular tecjniques of self-training (and its successor help-training) that are both meta-frameworks over regular classifier methods but require probabilistic representation of output, which can be hard to obtain in the case of discriminative classifiers. Therefore, we develop a new algorithm that uses the modified active learning technique Query-by-Committee (QbC) to sample data with high certainty from the unlabeled set and subsequently embed them into the original training set. Our new method allows us to achieve increased performance over both a range of datasets and a range of classifiers.udududThese three works are connected by gradually relaxing the constraints on the learning setting in which we operate. Although our main motivation behind the development was to increase performance in various real-world tasks (gesture recognition, sketch recognition), we formulated our work as general methods in such a way that they can be used outside a specific application setup, the only restriction being that the underlying data evolve over time. Each of these methods can successfully exist on its own. The best setting in which they can be used is a learning problem where the data evolve over time and it is possible to discretize the evolutionary process.ududOverall, this work represents a significant contribution to the area of both semi-supervised learning and pattern recognition. It presents new state-of-the-art techniques that overperform baseline solutions, and it opens up new possibilities for future research.