Model selection

What makes up a model?

And what choices do we have when building one? We must think about software ecosystem, size of model, task the model must accomplish, computational power of the end user device – we talk about these and more in this video.

We may wish to use transfer learning to “update” an existing model with our data. So, a model may have been “pre-trained” on a different, often larger dataset. We can take this model as a starting point, and just update some parts of the model so that it captures the nuances of our smaller dataset. There are different ways of doing this, but often with deep learning, certain parts or layers of the model architecture are “frozen” – we do not update them in the transferring stage. This makes updating even a large model to work with a smaller dataset a tractable problem.