Figure 4

Top-down multi-task (TDMT) training procedure. The example shows a taxonomy with two inner nodes and four leaves or tasks. A red task indicates that the instances of the task are used for model training, whereas a yellow task means that the instances are not used for training. For each node in the taxonomy a model is trained in a top-down fashion. (a) First, the root model is trained taking into account all training instances. (b) Next, the model of the inner node 2 is trained with the instances of the subtree. The model is required to be similar to the parent model ${\mathbf{w}}_p^{\ast }$ by the regularization term of Equation 15, which is indicated by a gray arrow. (c) Finally, the leaf model for task “T1” is trained using the instances of the task to compute the loss, while pulling the model towards the parent model. Procedure (c) is applied to all leaf nodes until we inferred a model for each task.