The current classifier uses a decision tree that creates a model for each species of bird. It performs with 0.48 on the leaderboard, despite multi changes to the efficacy of the embedding. I think the classifier is the bottleneck here -- it might be useful to build a classifier on top of the embedding using a simple dataloader that assumes that the category the audio file is part is an effective enough label.
The current classifier uses a decision tree that creates a model for each species of bird. It performs with 0.48 on the leaderboard, despite multi changes to the efficacy of the embedding. I think the classifier is the bottleneck here -- it might be useful to build a classifier on top of the embedding using a simple dataloader that assumes that the category the audio file is part is an effective enough label.
https://machinelearningmastery.com/multi-label-classification-with-deep-learning/
Are there ways to build a more effective no-call detector with the current dataset?