Deep Learning for Temporal Information Processing

Temporal information processing is required in a variety of learning problems where knowledge can be acquired from data of a sequential order or a stream, e.g. speech modelling, image sequence analysis, robot navigation, financial prediction and so on. Moreover, different learning paradigms such as supervised, unsupervised and reinforcement learning may be involved in temporal information processing. Recent studies suggested that deep learning has a great potential for temporal data analysis and information extraction regardless of learning paradigms but still has a number of weaknesses to be overcome in comparison to the existing statistical temporal information processing methods. In this fundamental research project, we are going to investigate several issues that hinder the existing deep learning models from working such as effective temporal information extraction and representations of complex temporal data, efficient learning and explainable and interpretable decision made by deep learning models. To address such issues, we are going to work on: 1) exploration of emerging deep architectures and learning algorithms to overcome the weakness of the existing deep learning techniques for effective extracting and encoding temporal information; 2) development of efficient learning strategies for novel deep architecture via transfer learning; and 3) Improvement of explainablity and interpretability of novel deep learning models via feature importance estimate. To demonstrate the effectiveness of our proposed methods, we are going to apply our techniques to deal with challenging time series forecast problems, such as those posed in M4/M5 forecasting competitions. This project is closely related to EPSRC strategy focus: Artificial Intelligence Technologies.