Unsupervised Continual Learning on Large Scale Data Streams

This project falls within the EPSRC Information and communication technologies research area. Social networks provide an environment where anyone can share their beliefs and experiences with the world. Popular social networks see huge numbers of new messages and interactions posted every minute. Unfortunately, messages and interactions involve content that be regarded as an online harm, e.g., mis/disinformation. Misinformation thrives in online communities that lack systematic moderation, allowing social media 'influencers', bots and users, to amplify the spread of online hate or misinformation. Current Artificial Intelligence (AI) and Machine Learning (ML) approaches have significant drawbacks which means they are often not the practical solution to problem; this PhD aims to address this.

The first is that current AI and ML methods are typically offline learners, trained with a static dataset and usually with a set of labels. However, in many applications, such as the one noted, data is constantly being created, updated and new 'labels' are dynamically created. This means that models trained offline are unsuited for a practical implementation as they do not perform the task they have been created for. In addition, these models cannot improve when they are in use as they need a supervised data stream in order to update. This is clearly at odds with current systems, where the purpose of classification implies that a manual content-analysis is too expensive (in time or money). Therefore, future models need to be able to adapt to a continuously changing unlabelled data stream.

The second is that these systems typically ignore the multimodal nature of the world in which they operate. On social networks, for example, the data consists of text, images, videos and interactions. To be most effective, AI and ML systems should learn and make decisions considering all forms of data modalities. The effectiveness of multi-modal algorithms is currently relying heavily on the knowledge of domain experts to design architecture that is specialised for the task. This means that it is hard to generalise multi-modal models as implementation's are very much designed with the task at hand in mind.

Our novelty is to satisfy these constraints in order to build the next generation of AI systems that are able to tackle extremely complex real-world challenges across many domains and applications.

This includes, but is not limited to, applications involving tasks on sensor networks, the world wide web, social networks and any generic stream of data where something must be learned, and decisions must be made with little apriori knowledge. Importantly, contributions from this work (such as the continual learning components) can be used in many, if not most, areas where current machine learning is used and static training sets are not as effective after the models are moved to production due to problems such as the emergence of new classes, concept drifting and distribution changes.