PROJECT TITLE :

Exploiting Related and Unrelated Tasks for Hierarchical Metric Learning and Image Classification

ABSTRACT:

Multiple connected activities are taught together in order to improve performance in multi-task learning. While it is possible to detect which tasks are connected, it is also possible to determine which tasks are unconnected. When it comes to multitask learning, researchers have traditionally focused on leveraging correlations between related tasks rather than prior knowledge from unrelated tasks. Hierarchical multi-task metrics can be learned by exploiting existing knowledge about both related and unrelated activities, as demonstrated in this study. In the first step, a visual tree is built to hierarchically organise a huge number of image categories in a coarse-to-fine form. To learn a multi-task classifier for each node in the visual tree, we use both related and unrelated tasks, where the learning tasks for sibling child nodes under the same parent node are viewed as correlated tasks, and the other learning tasks are considered unrelated. It also propagates the parent node's node-specific metrics to its sister child nodes to regulate error propagation between levels. Hierarchical metric learning delivers superior results than current state-of-the-art methods, according to our testing results