Training machines to multi-task
Machine learning is the study of computer programmes that allow computers (or machines) to learn to do things and to improve their own performance. It is an artificial intelligence approach that is widely used to classify data records into discrete categories or labels. However, in many tasks, some labelling decisions are inter-related and have to be decided simultaneously. Such complex tasks are called structured predictions. There are problems such as error propagation with the current method of dealing with these complex tasks. The goal of the EU-funded Jointstructuredpred project was to overcome these limitations. The project proposed a method for dealing with complex tasks that involved training the subtasks jointly using multi-task learning techniques. To compare this new approach and the current one, a software programme was developed for both methods for structured prediction. An evaluation of both methods using typical problems in computational biology showed that the proposed method outperformed the current method by significantly improving the prediction accuracy. The findings were similar, but not as pronounced, when they were compared using typical applications in natural language processing. Natural language processing is the computerised approach to processing human language in terms of its meaning while computational biology combines computer science and molecular biology. Project results indicate that the proposed approach can be used to improve the performance of many prediction problems in a wide range of disciplines, including biology, medicine, linguistics and signal processing.
