Few labels, shifting data

Make use of unlabeled data, and of data that does not look like the training data

3 exercises

Each exercise is graded automatically, in your browser. The full statements are available once you have an account.

1. What this is for in real life

   Labelling data is expensive, and the real world never quite looks like the training data. - Sound ecology. Automatic recorders capture thousands of hours of songs in forests and marshes. Tools such as BirdNET…

2. The sound archive

   You have just joined a bird observatory. All spring, its automatic recorders captured birdsong, and each recording is summarized by eight acoustic descriptors: peak frequency (kHz), bandwidth (kHz), duration (s), number…

3. The new listening sites

   Thanks to your work, the forest archive is now fully identified: 1,500 songs, with their species (X_train, y_train, the same eight descriptors, six species numbered 0 to 5). The station is now installing its…