Skip to main content

Multi-Task Learning

The same network can usually do different things, we just need to change the output shape and/or loss.

Pros and Cons​

Pros

  • Usually has a low overhead
    • Can just append a couple of extra layers to get another output
    • Cheaper than having a network for each task
  • Usually helps learning
    • One task helps to regularize the other
    • Particularly if tasks are related

Cons

  • We now need two sets of labels

Loss Weights​

We can change the loss weights to help the network consider all tasks differently.

This can be achieved by providing a list or dict of loss weights to the compile command when we build the model.

model_cnn.compile(loss=['mean_squared_error', keras.losses.SparseCategoricalCrossentropy(from_logits=True)],
                  loss_weights=[1, 100],