Multitask/Multioutput QEPs with Exact Inference

Exact QEPs can be used to model vector valued functions, or functions that represent multiple tasks. There are several different cases:

Multi-output (vector valued functions)

• Correlated output dimensions: this is the most common use case. See the Multitask QEP Regression example, which implements the inference strategy defined in Bonilla et al., 2008.

• Independent output dimensions: here we will use an uncorrelated QEP for each output.

  • If the outputs share the same kernel and mean, you can train a Batch Uncorrelated Multioutput QEP.

  • Otherwise, you can train a ModelList Multioutput QEP.

Scalar function with multiple tasks

See the Hadamard Multitask QEP Regression example. This setting should be used only when each input corresponds to a single task.