BirdNET: workflow for correcting manually trained classifier?

Question

I am using BirdNET GUI for analysis of bird vocalizations. I have created a classifier using the GUI and some training data, and I am wanting to improve it, e.g. correct false classifications and allow the classifier to learn from this. Does anyone know of a workflow for this, either within the GUI or using RavenPro or other audio analysis software? I know it is a relatively new platform, and I haven't been able to find reference materials about this process.

Answer 1

I don't think the BirdNET GUI contains anything for "correcting" or re-training the classifier.

However, using the BirdNET command-line version you can train the algorithm yourself - instructions here - and this is what you want to do. In this context, "correcting" the algorithm would be re-training it with the original training dataset (ideally) PLUS the examples that you've manually labelled that the algorithm was getting wrong.

The "original training dataset" might be unavailable or too big for you to use for re-training. Perhaps try making a not-too-small generic dataset of all the species you're interested in, plus the manual corrections.

Answer 2

I agree with Dan, I'll just mention another possible direction:

An iterative training workflow, known as "active learning" or "human-in-the-loop", takes a bit more work to set up than just training a custom classifier on top of BirdNET. Researchers at Google have provided some tutorials and notebooks for active learning (they also call it "agile modeling") relying on their pre-trained model Perch, rather than BirdNET: https://colab.research.google.com/drive/1gPBu2fyw6aoT-zxXFk15I2GObfMRNHUq?usp=sharing

It will require more hands-on programming than the BirdNET interface.

Answer 3

I have been doing the same over te past months. Generating a classifier with training data and then regenerating a new classifier with the false positive data of the first classifier. I just add them to a folder "Noise" in the folder of the training data and then train a new classifier. This has slowly been working, but is an incredible iterative process that takes a lot of time for me unfortunately.