Week 3 review

We hope it has given you some pointers on what to think about when selecting and training a machine learning model using your own experimental data.

We discussed the following topics:

• software environments – e.g. Pytorch, Tensorflow, Scikit-Learn
• model selection – what architecture to use
• transfer learning – using pre-trained model weights from another model
• training versus inference – inference is using the model on fresh data once trained
• troubleshooting and debugging
  • problems that produce runtime errors
  • problems which don’t produce errors, but the model still fails to train properly
• parameter tuning – adjusting hyperparameters
• multitask learning – more than one cost function and objective to the model, e.g. counting leaves and identifying a species name.

What’s next?

By this point we might have a machine learning model that we think performs pretty well on our chosen task. But what else do we need to consider? How well can we trust our results? What do they mean? And how can we improve them? We will consider these questions in the next week of the course.