How Do Machine Learning Models Improve Over Time?

The main question following 'Machine Learning for Humans' talk in Cambridge

  • Data Science
  • Machine Learning
  • Presentations
Posted on

How Do Machine Learning Models Improve Over Time?

The main question following 'Machine Learning for Humans' talk in Cambridge

Peter Marriott

Last week I presented a talk on ‘Machine Learning for Humans’ at Cambridge DDD Nights (the slides are available together with the talk abstract is on my original post

Following my talk the burning question was about how these machine learning models improve over time.

Data is the Magic

Machine learning doesn’t mean that a computer can magically improve the model on it’s own. However, just as we can improve our abilities by receiving feedback, we can arrange for the machine learning models to improve by the use of the same mechanism.

In the case of reinforcement learning the feedback is built-in; the whole way this sort of learning works is by receiving a success/failure response or score for each output during the training phase.

For the more traditional supervised learning we can add or tune the training data. After this has been done we look at the metrics on the test data set to see whether the changes to the training data have improved the performance. If the test data set is representative of your actual data you should see improvements in the live situation.