Parameter tuning and multitask learning

Performance ok but could be better?

What if the model is producing a decent result, but it could be much better? The fundamentals of the model and training setup might be fine, but perhaps we can improve our predictions by fine tuning some parameters. Many pipelines have “hyperparameters”, parameters which describe the training process itself, for example, rather than the internal weights of the models. Any example of this is the “learning rate” during training, which is discussed in the video.

Other options to improve performance are multi-task learning – asking a model to do two different tasks at once. As we see in the video, sometimes this can improve the overall performance of a model.