Week 4 Review

In the videos we talked about various different aspects of the outputs and outcomes of machine learning and deep learning models that you may need to consider. We hope the videos give you some ideas of the questions you may need to be asking of both your own models, and the work of other researchers.

In summary, we considered the following topics:

• trusting results
  • what results could have been obtained purely at random?
  • how accurate do your results need to be? – depends on what the model is to be used for
  • use of other metrics – e.g. precision and recall
• interpreting output
  • looking at mistakes – is there is common problem?
  • how certain are model decisions? – generally a probability score
  • does it out-perform a human?
  • what areas of an image is the network ‘looking’ at? – Cumulative activation maps
  • is explainability important or not?
• improving results
  • what happens after a model is released or published?
  • portability of code – can other researchers use your code?
  • can other researchers improve your model?
  • ethical aspects – how will you and others use the training data and models?

What’s next?

In the final week of the course we will offer some tips and tricks on how to put all we have learned so far together, including software and hardware tips, writing academic papers containing machine learning, and engaging with researchers within the wider computer science discipline.